JP4150554B2 - Object image extraction method, object image extraction program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an object image extraction method and an object image extraction program for obtaining an object region in an image.
[0002]
[Prior art]
A technique for extracting a region of an arbitrary object from an image is useful when processing an image, for example, replacing a background with another image. For this purpose, it is necessary to correctly determine the region of the object along its contour.
[0003]
Specifically, the object region is represented by an image called an alpha mask (or shape data). An alpha mask is an image in which different pixel values are set in an area where an object exists (object area) and other areas. An example of the original picture and its alpha mask are shown in FIGS. 3 and 4, respectively. FIG. 4 shows the alpha mask of the right ship 031 shown in the original picture of FIG.
[0004]
In other words, extracting an object is creating an alpha mask. Once the alpha mask is obtained, the background area other than the object area can be easily replaced with another image by referring to it.
[0005]
There is a need for a technique for obtaining an accurate shape of an object based on the approximate shape when the approximate shape of the object is tentatively obtained in advance by some method during such object extraction. As one of the methods, an object extraction method (fractal method) using a self-similar map has been proposed (see, for example, Patent Document 1, Patent Document 2, and Non-Patent Document 1). The outline of this technique will be described with reference to FIGS. 5, 6, 7, and 8.
[0006]
FIG. 5 shows a state in which a schematic shape 051 is given to the ship 031, and its alpha mask is FIG. 6. At this stage, the contour of the alpha mask is shifted from the contour of the object. In the fractal method, for such input of original image data and an alpha mask, first, a plurality of square blocks are set along the general shape of the object using the alpha mask. These blocks are called child blocks. Then, a child block is set at the same position in the original image data as in the alpha mask. FIG. 7 shows a child block set in the original image data. For easy understanding, the rough shape is also shown in the original image data.
[0007]
Next, a parent block corresponding to each child block is obtained using the original image data. The size of the parent block is twice the height and width of the child block, and the image pattern (design) in the parent block is the same as the image pattern in the child block.
[0008]
In order to find a parent block similar to the child block pattern, a double-vertical block is set as a candidate block near the child block on the original image data. Then, the error between the reduced block and the child block obtained by reducing the candidate block to 1/2 in the vertical and horizontal directions is obtained.
[0009]
In the case of a grayscale image, for example, an absolute value error (the absolute value of the difference between each pixel is added to all the pixels in the block) is used as the error. Such error evaluation is performed on several candidate blocks, and the candidate block having the smallest error from the child block is determined as the parent block. Such parent blocks are obtained for all the child blocks shown in FIG.
[0010]
FIG. 8 is an example of the parent block 082 obtained for the child block 081 set at the tip of the ship 031. It can be seen that the symbols in each block are similar. Although a binary image is shown in FIG. 8, the same applies to a grayscale image in which 256-level pixel values are given to pixels.
[0011]
With the above processing, data of a child block and a parent block with a similar pattern can be obtained. Next, set the parent block at the same position of the alpha mask as the original image data. To do.
[0012]
The object extraction is performed by correcting the alpha mask using the child block and the parent block. Specifically, the alpha mask data of each child block may be replaced with the alpha block data of the parent block corresponding to each child block reduced to 1/2 in the vertical and horizontal directions.
[0013]
When this replacement conversion (self-similar mapping) is performed once, the approximate shape slightly approaches the correct shape. The same replacement conversion is performed again on the approaching alpha mask. In this way, when the conversion is recursively replaced with the alpha mask, it converges to a substantially correct shape. An alpha mask that has been subjected to replacement conversion a certain number of times is taken as the object extraction result.
[0014]
[Patent Document 1]
JP 2000-82145 A
[0015]
[Patent Document 2]
JP 2001-188910 A
[0016]
[Non-Patent Document 1]
Ida, Sanbonsugi, “Contour Fitting by Self-Similar Mapping”, Proceedings of the 5th Image Sensing Symposium, C-15, pp. 115-120, June 1999.
[0017]
[Problems to be solved by the invention]
However, in the conventional fractal method, when another object is nearby, the contour of the other object may be extracted. In addition, when the object shape is thin, the object shape may disappear, and the object shape may not always be extracted accurately.
[0018]
Therefore, an object of the present invention is to provide a method and program that can accurately extract the shape of an object even when another object is nearby or the object shape is thin.
[0019]
[Means for Solving the Problems]
In order to solve the above-mentioned problem, the object image extraction method according to the present invention includes a first step of receiving image data including a target object to be extracted and shape data representing a region of a schematic shape of the target object in the image data. A second step of setting a plurality of search reference blocks by shifting their positions along the outline of the schematic shape represented by the shape data, and each search reference block being larger than the search reference block A third step of searching for a similar block in which the design of the area corresponding to this block on the image data is most similar to the design of the area corresponding to the search reference block on the image data; The shape data of the region corresponding to each search reference block is changed to the shape data of the region corresponding to each similar block. A fourth step of repeatedly performing the replacement with the reduced one until a predetermined condition is satisfied, and the first step represents the certainty that the target object exists in each pixel in the image data The label data is also received, and in the third step, a weighted sum of errors of the entire block is calculated by referring to the label data and ignoring or neglecting an error between pixels that are unlikely to have a target object. A search is performed for a similar block that minimizes the weighted sum.
[0020]
In addition, the shape data represents a region of a schematic shape of the target object, and each pixel thereof has a one-to-one correspondence with each pixel of the image data, and the label data includes the image data as the target object. It is assumed that the region is divided into at least two types out of three types: a fixed region, a non-target object fixed region, and an undetermined region, and each pixel corresponds to each pixel of the image data on a one-to-one basis. In step 3, for each of a plurality of candidate blocks that are candidates for the similar block, data in a region corresponding to the search reference block in the image data, and data obtained by reducing a region corresponding to the candidate block in the image data; The error of each corresponding point is calculated, and the weighted sum is calculated in the region where both of the corresponding points are classified into the target object determination region by the label data. If both are present in a region that is classified as a non-target object determination region, the error about the corresponding point is ignored or neglected, and the candidate block that minimizes the weighted sum of the error is It is good also as a feature to determine as a similar block.
[0021]
In the third step, for each of a plurality of candidate blocks that are candidates for the similar block, data in a region corresponding to the search reference block in the image data and a region corresponding to the candidate block in the image data are obtained. The error of each corresponding point with the reduced data is calculated, and the weighted sum is calculated when both the corresponding points are present in the region classified as the target object determination region by the label data, or both are non-target objects. If it exists in a region classified as a definite region, a weighting factor to be multiplied by the error for the corresponding point is determined as 0 or a small value, and a candidate block with the smallest weighted sum of the errors is determined as the similar block It may be characterized by.
[0022]
The object image extraction method of the present invention includes a first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data, and the shape data. A plurality of search reference blocks are set so as to be shifted from each other along the outline of the outline shape, and a pattern of an area corresponding to the search reference block on the shape data is thin of the target object. The pattern of the portion is a second step of canceling the setting, and each search reference block is larger than the search reference block, and the design of the area corresponding to that block on the image data is the search reference block A third step of searching for a similar block most similar to the pattern of the region corresponding to, and in the shape data Yes serial processing for replacing the data of the region corresponding to each search criterion blocks obtained by reducing the data of the region corresponding to each of the similar block, and a fourth step of performing repeated until a predetermined condition is satisfied In the second step, the number of changes in the pixel value is obtained by scanning the circumference of the region corresponding to the search reference block on the shape data, and when the number of changes exceeds two, the shape data It is determined that the pattern of the area corresponding to the search reference block above is a pattern of a thin portion of the target object. It is characterized by that.
[0024]
The object image extraction method of the present invention includes a first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data, and the shape data. A second step of setting a plurality of search reference blocks to be shifted from each other along the outline of the approximate shape, and each search reference block is larger than the search reference block and is on the image data A similar block whose pattern in the area corresponding to the block is most similar to the pattern in the area corresponding to the search reference block is searched, and the pattern in the corresponding area on the shape data is a pattern of a thin portion of the target object. A third step of canceling the setting of the similar block and the search reference block corresponding thereto, and the shape data Yes the processing for replacing the data of the region corresponding to each search criterion blocks obtained by reducing the data of the region corresponding to each of the similar block in data, and a fourth step of performing repeated until a predetermined condition is satisfied In the third step, the number of changes in the pixel value is obtained by scanning the circumference of the area on the shape data corresponding to the similarity block. If the number of changes exceeds 2, the similarity is determined. Determining that the pattern of the area on the shape data corresponding to the block is a pattern of a thin portion of the target object; It is characterized by that.
[0026]
The object image extraction method of the present invention includes a first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data, and the shape data. A second step of setting a plurality of search reference blocks to be shifted from each other along the outline of the approximate shape, and each search reference block is smaller than the search reference block and on the image data A third step of searching for a similar block whose pattern corresponding to the block most closely resembles the pattern of the area corresponding to the search reference block; and data of areas corresponding to the similar blocks in the shape data, respectively The process of replacing the data in the area corresponding to the search reference block with the reduced data satisfies the predetermined condition. Characterized by a fourth step of performing repeatedly.
[0027]
The object image extraction method of the present invention receives image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data, and expands these to expand image data and an enlarged image data. A first step of generating shape data; a second step of setting a plurality of search reference blocks by shifting their positions along the outline of a schematic shape represented by the enlarged shape data; and the search For each reference block, a third block is searched for a similar block which is larger than the search reference block and whose symbol on the enlarged image data is most similar to the symbol of the region corresponding to the search reference block. Step, and the data of the area corresponding to each similar block in the enlarged shape data to each search reference block The fourth step of repeatedly replacing the data in the corresponding area with the reduced data until a predetermined condition is satisfied, and the fifth step of reducing the enlarged shape data obtained in the fourth step to the original size And a step.
[0036]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, a plurality of embodiments of the present invention will be described with reference to the drawings.
[0037]
(Items Common to Each Embodiment) FIG. 40 is a diagram of a PC (personal computer), WS (workstation), etc. used in each embodiment of the present invention. The PC and WS used in each embodiment include an input device 4007 such as a pointing device such as a mouse, a keyboard, and other external devices (for example, a digital camera), an input reception unit 4004 that receives input from the input device 4007, a CRT, and an LCD. And an image output unit 4006 for outputting an image signal thereto.
[0038]
The PC and WS used in each embodiment include a magnetic disk drive 4002 such as a hard disk, a memory 4005, a central processing unit 4001, and an optical disk drive 4003.
[0039]
Each of the embodiments of the present invention is an object image extraction program that runs on a PC (personal computer) or WS (workstation). The user reads a program stored in advance in the magnetic disk drive 4002 into the memory 4005 and causes the PC or WS to execute it.
[0040]
Before executing the object image extraction program, the user acquires image data to be subjected to object shape extraction processing and stores it in the magnetic disk drive 4002 in advance. At the time of execution, the object image extraction program reads this image data from the magnetic disk drive 4002 and stores it in the memory 4005.
[0041]
Then, the image data is displayed on the display device 4008, and the user is requested to input the alpha mask data of the approximate shape specifying the approximate shape of the object by using the mouse. The input alpha mask data is also stored in the memory 4005.
[0042]
Further, the object image extraction program extracts the object shape based on the image data and the alpha mask data of the approximate shape. By extracting the object shape, an alpha mask having an accurate shape is generated and output as an extraction result.
[0043]
The alpha mask of the contour extraction result is stored in the magnetic disk drive 4002 or the memory 4005 provided in the PC or WS. For example, it is used in another program or apparatus that cuts out an image of an object from an image.
[0044]
Note that a camera may be connected to a PC or WS so that the user can acquire an image at an arbitrary timing. Further, the data of the substantially shaped alpha mask may be automatically generated using a background difference method or the like.
[0045]
In the following description, the search reference block is referred to as a child block, and a similar block that is larger than the search reference block (for example, twice in the vertical and horizontal directions) and similar in design is referred to as a parent block.
[0046]
(First Embodiment) This embodiment is to improve that the object shape may not be accurately extracted when there is an edge such as another object contour in the vicinity of the contour extraction target object. .
[0047]
Hereinafter, the processing of this embodiment will be described with reference to the flowchart of FIG.
[0048]
(Step 010) Image data, a substantially shaped alpha mask, and label data are prepared on a memory.
[0049]
As the image data, for example, image data that has been previously captured by a user with a digital camera or the like and stored in a hard disk is read out.
[0050]
The alpha mask is data representing the shape of an object created by the user specifying, for example, a mouse while viewing the image data displayed on the image display device. The alpha mask is data having the same number of pixels as the image data. The pixel value is 1 when the object is present and the pixel value is 0 when the object is not present.
[0051]
The label data is data representing information related to an object area and a background area different from the alpha mask, and is created by a user using a mouse while viewing image data. For each pixel, for example, a region that is known to be an object other than the background or the object in advance is a non-target object determination region (background determination region), and the pixel value of each pixel is 0, and the region is known to be an object in advance. , The pixel value of each pixel is set to 1 as the target object fixed region (object fixed region), and the pixel value of each pixel is set to 2 as the undefined region for the background or the object. The label data may be automatically generated with reference to, for example, an alpha mask.
[0052]
(Step 011) The child block C is set in the contour portion of the alpha mask. In the present embodiment, the child blocks C are set so as to overlap each other as shown in FIG. 7, but may be set so as to contact each other without overlapping. The position information (for example, the coordinates of the upper left corner and the coordinates of the lower right corner) of the set child block C is stored in the memory.
[0053]
(Step 012) A parent block P which is a block similar to each child block is found with reference to label data and image data.
[0054]
First, a candidate block that is a parent block candidate is set in the vicinity of the child block. Then, on the image data, the error is obtained by referring to the label data of the child block region pattern and the candidate block region pattern. An error is calculated for a plurality of candidate blocks, and a block having the smallest error is defined as a parent block. Since the candidate block is a block having the same number of pixels as the parent block, when the error between the child block and the candidate block is obtained, the candidate block is reduced to have the same number of pixels as the child block.
[0055]
The error calculation with reference to the label data is performed as follows. For example, when the user designates the area inside the area 151 in FIG. 15 as an area in which the object is not extracted, the label data has a pixel value of 0 inside (background determined area), and the other pixel values of 2 (undefined area) ). And the pixel with the child block is X _c , The pixel of the corresponding reduced block is X _p X _c And X _p Is defined as follows. (1) X _c And X _p When both are within the background determination area, _c And X _p “0” is ignored.
(2) X _c And X _p When both are within the object determination area, _c And X _p “0” is ignored.
(3) If neither (1) nor (2) above, _c And X _p Error "=" X _c And X _p The absolute value of the difference between the pixel values of ".
[0056]
X obtained in this way _c And X _p Are added to each other in the block to obtain an error between the candidate block and the child block. Then, the candidate block with the smallest error is determined as the parent block. Then, the determined parent block position information is stored in the memory in association with the child block.
[0057]
(Step 013) Step 012 is repeated until parent blocks for all child blocks are determined.
(Step 014) On the alpha mask, the data of the area corresponding to each child block is replaced with the data of the parent block corresponding to the child block reduced to the same number of pixels as the child block.
(Step 015) Step 014 is performed for all child blocks. As a result, all child blocks of the alpha mask are replaced once.
(Step 016) Steps 014 and 015 are recursively repeated a predetermined number of times. Thereby, the contour of the alpha mask asymptotically approaches the contour of the object.
[0058]
In the error calculation in step 12, in the cases (1) and (2), the error is ignored and set to 0. However, even if the error is neglected, that is, multiplied by a predetermined coefficient larger than 0 and smaller than 1, good. Alternatively, in the cases (1) and (2), the calculation itself may be omitted ignoring the error.
[0059]
Hereinafter, an example of extracting an object shape by a conventional fractal method and an example of extracting an object shape according to the present embodiment will be compared.
[0060]
An example will be described in which an object shape is extracted using a conventional fractal method when another object is nearby. The child block 091 shown in FIG. 9 includes the area of the upper left ship 032 in addition to the area of the lower right ship 031. In this case, there is a high possibility that a candidate block that is completely similar to the child block 091, that is, the absolute value error is 0, is not found, but the candidate block with the smallest error is selected as the parent block.
[0061]
In order to extract the ship 031, it is convenient to select a block 093 similar to the area of the ship 031 as the parent block. This is because the contour of the alpha mask converges to the contour of the ship 031 by the replacement conversion. However, the block 093 is not necessarily selected.
[0062]
FIG. 11A is an enlarged view of the periphery of the child block 091, and shows a parent block candidate block 093, a contour 114 of the ship 032, and a contour 115 of the ship 031. In FIG. 11A, parent block candidate block 093 is set to be similar with respect to contour 115 of ship 031.
[0063]
FIG. 11B shows an image pattern 111 of the child block 091, an image pattern 112 obtained by reducing the block 093, and a difference pattern 113 between the image pattern 111 and the image pattern 112.
[0064]
Here, the pixel values of the area of the ship 032 and the area of the ship 031 painted with horizontal lines are set to the same value, and the other areas are set to be constant with different pixel values. In the difference pattern 113, a portion where an error in pixel values occurs in the image pattern 111 and the image pattern 112 is painted black. The absolute value error is proportional to the area of this black region.
[0065]
On the other hand, FIGS. 12A and 12B show a case where the block 121 which is a parent block candidate is set to be similar with respect to the contour 114. In this case, the area of the black region of the difference pattern 122 is smaller than that of the difference pattern 113 in FIG. As a result, as shown in FIG. 10, the block 121 is selected as the parent block, and the contour of the alpha mask of the ship 031 approaches the ship 032 due to the replacement conversion for this portion, so that a correct contour cannot be obtained.
[0066]
On the other hand, a parent block convenient for object extraction is selected by processing using label data as in step 012. This will be described using an example in which an alpha mask as shown in FIG. 4 is generated from the original image shown in FIG.
[0067]
FIG. 5 shows a state in which a rough shape of an object is given, and FIG. 6 shows an alpha mask of this rough shape. FIG. 7 shows the result of arranging child blocks.
[0068]
FIG. 13A is an enlarged view of the periphery of the child block 091. FIG. The contour 114 is the contour of the ship 032, and the contour 115 is the contour of the ship 031. The parent block 093 is arranged so as to be similar with respect to the ship 031 outline 115. The upper left side of the background determination area line 131 that is the boundary line of the background determination area is the background determination area 151 shown in FIG.
[0069]
FIG. 13B shows an image pattern 132 of the child block 091, an image pattern 133 of the parent block 093 (reduced to the same size as the child block), and a difference pattern 134 between the image pattern 132 and the image pattern 133. It is a figure.
[0070]
In FIG. 13B, the image pattern 132 and the image pattern 133 also show a background determination area line 131, and the difference pattern 134 includes a common area between the background determination area in the image pattern 132 and the background determination area in the image pattern 133. The boundary line 135 is also shown.
[0071]
As described with reference to FIGS. 11A and 11B, in the conventional arrangement of such parent blocks, many errors have occurred on the upper left side. However, in the present embodiment, in both the image pattern 132 and the image pattern 133, the error of the pixel that is the background determination region is zero, so that no error appears in the region on the upper left side of the boundary line 135 of the difference pattern 134. . Therefore, the error in the entire block is reduced as in the difference pattern 134 as compared with the conventional case.
[0072]
FIGS. 14A and 14B are diagrams in the case where the parent blocks are arranged so as to be similar with respect to the contour 114 of the ship 032 as in FIGS. 12A and 12B. In this case as well, the error on the upper left side of the line 135 is canceled as in the case of FIG. 13. However, since there is no error on the upper left side of the line 135, the error on the lower right side as a whole is the same as in the case of FIG. An error occurs.
[0073]
Comparing FIG. 13 and FIG. 14, in this embodiment, the case of FIG. 13 has a smaller error than the case of FIG. 14, and the parent block so as to be self-similar with respect to the case of FIG. Is selected, the contour of the ship 031 is correctly extracted.
[0074]
In the present embodiment, the label data is specified in advance by the user using the mouse, but the label data may be obtained by using a method for automatically obtaining the label data described below.
[0075]
In this embodiment, step 014 and step 015 are repeated a predetermined number of times, but the condition for repetition is not limited to “number of times”. For example, when the replacement is performed in step 015, the sum of the absolute value difference between the alpha mask of the child block and the reduced alpha mask of the parent block is calculated, and the calculation is repeated until the sum of the absolute value differences becomes a predetermined threshold value or less. You may do it.
[0076]
(First Label Data Generation Method) For example, when the color of an object or the color of a background is known in advance, there is also a method in which the color areas are set as an object determination area and a background determination area, respectively.
[0077]
This method will be described by taking as an example the case where the schematic shape 051 is given as described above.
[0078]
(1) The shape 371 is obtained by expanding the schematic shape 051 with a constant width W as shown in FIG. Similarly, the shape 372 is obtained by inflating with a width of 2 W.
[0079]
Here, dilation is a type of processing by a morphological filter. When a pixel other than the processing target region (here, the object region) is scanned and the processing target region is within a certain distance from the target pixel, that pixel is used. Is a process of adding to the processing target area.
[0080]
(2) The general shape 051 is contracted by a width of W and 2W to obtain a shape 373 and a shape 374, respectively.
[0081]
Here, contraction is also a type of processing by the morphological filter. When there is a region other than the processing target region within a certain distance from the target pixel while scanning the pixel of the processing target region, processing for removing the pixel from the processing target region It is.
[0082]
(3) A region 381 sandwiched between the shape 371 and the shape 372 (a region indicated by hatching in FIG. 38) is set as a background color sampling region. In addition, a region 382 (region indicated by a horizontal line in FIG. 38) sandwiched between the shape 373 and the shape 374 is a foreground color sampling region. Here, the reason why the colors are sampled from the portion away from the approximate shape 051 by a certain distance is that the approximate shape is deviated from the correct shape.
[0083]
(4) A background color table is created by extracting all the colors of the pixels in the background color sampling area in the original image. Similarly, a foreground color table is created.
[0084]
Instead of extracting all the colors, when a certain number of pixels are the same color, the colors may be registered.
[0085]
(5) The background color table and the foreground color table are compared, and if the same color is registered, the color is removed from both tables.
[0086]
(6) Each pixel of the original image is scanned using the two color tables obtained in this way. Then, label data is created by setting the color pixel in the background color table as the background determination area, the color pixel in the foreground color table as the object determination area, and the color pixel not in any table as the undetermined area.
[0087]
By doing as said (1)-(6), the area | region which can be estimated with a background or an object from a viewpoint of a color can be made into each definite area | region.
[0088]
Rather than registering similar colors individually, if the color level is created beforehand and the label data is created after the color level of the original image is quantized in advance, the number of colors handled will be reduced and the amount of processing can be reduced. Contributes to faster processing.
[0089]
(Second Label Data Generation Method) Locally optimized information can be obtained by creating a color table locally on the screen and creating label data while switching each.
[0090]
As an example of creating a color table locally, the following method can be considered. For example, as shown in FIG. 39, blocks 391, 392 and the like are arranged along a schematic shape 051, and the above-described color table is created only inside each block. Then, when creating the label data, it is determined whether the background is determined, the object is determined, or is not yet determined using the color table of the block to which each pixel belongs. At that time, in the portion where the blocks 391 and 392 overlap, first, a label is determined by the color table of one of the blocks, and if the label is undecided, the label is reattached by the other color table.
[0091]
(Third Label Data Generation Method) When generating a background color table or an object color table, a method of registering and generating the color of a pixel that indicates a screen as a background or object instead of using a morphological filter is also available. is there.
[0092]
(Third label data generation method) When a background image in which an object to be contour extracted is not captured can be acquired in advance, a difference between the background image and the current image is obtained, and pixels with a difference equal to or greater than a certain value are detected. There is also a method of confirming.
[0093]
(Fourth Label Data Generation Method) When it can be assumed that there is a correct contour within a certain range from the contour of the general shape, the result of contracting the background region of the general shape is used as the background determination region, or the object of the general shape The result of contracting the region can also be used as the object determination region. With this method, the periphery of the processing target area is reduced.
[0094]
(Fifth label data generation method) When it is known that the object area of the approximate shape always includes the entire object, the background area is determined by using the background area of the approximate shape as it is or by contracting one or two pixels. This is an area. On the other hand, an object region having a substantially shaped shape is defined as an object determination region by contracting several pixels, for example, about ½ of the number of pixels on one side of a child block.
[0095]
Furthermore, the object determination area and the background determination area determined in this way may be expanded using the area growth method. On the other hand, the same applies to the case where the entire object region having a substantially shape is always included in the object.
[0096]
Here, the area growth method compares the pixel values inside and outside the area at the boundary of the growth target area, and if the difference between the pixel values is less than a certain value, the outside pixel is taken into the area. It is to go. By repeating this process, a portion having a close pixel value is taken into the region.
[0097]
(Sixth Label Data Generation Method) When one point on the object is input by the user, the object region may be obtained by the region growth method using that point as a starting point, and may be used as the object determination region. Similarly, for the background, the background definite region can be set using the region growing method.
[0098]
(Number of stages of label data) In the label data generation method up to this point, an example has been described in which background determination, object determination, and undecided three-stage data are generated, but the number of stages may be increased. For example, the pixel value of the label data is brought closer to 0 as the probability of being the background is higher, the pixel value is brought closer to 255 as the probability of being the object is higher, and the undetermined region is considered to correspond to the pixel value 128.
[0099]
The background and object colors are examined in some way (for example, by the background color sampling and object color sampling methods in the first label data generation method described above). Then, label data is generated by determining the pixel value of the label data according to whether each pixel is closer to the background or the color of the object.
[0100]
In the parent block search using the label data generated in this way, the pixel value of a pixel having a child block is set to X _c , The pixel value of the reduced block pixel of the parent block candidate is X _p , "X _c And X _p The absolute value of the difference between the pixel values of E is E, and the weighting coefficient for error calculation is W.
"X _c And X _p Error "= W x E
Is calculated for each pixel to obtain the sum, and “X _c And X _p The one with the smallest error is adopted as the parent block.
[0101]
Here, the weighting factor W is
(A) X _c And X _p When the pixel values of both of the label data are not close to 0 or 255, they are always kept at 1.
(B) X _c Or X _p When the pixel value of any one of the label data is close to 0, the weighting factor W is made smaller than 1 as the other label is smaller.
(C) X _c Or X _p When the pixel value of any one of the label data is close to 255, the weighting factor W is made smaller than 1 as the other label is larger.
[0102]
By controlling W in this way, it is possible to reduce the influence of errors between pixels having a high probability of being an object and pixels having a high probability of being a background in multiple stages.
[0103]
As described above, in the present embodiment, by using the masking information called label data when extracting the contour of the object, the contour can be accurately extracted even when a plurality of objects are close to each other.
[0104]
(Second Embodiment) A second embodiment of the present invention will be described below with reference to FIGS. This embodiment is an embodiment that is effective when a plurality of objects are extracted sequentially one after another. Here, the description will focus on the parts different from the first embodiment.
[0105]
In the image shown in FIG. 3, there are a ship 032 and a ship 031. In this embodiment, these contours are sequentially extracted. Here, first, the ship 032 is extracted by the method described in the first embodiment, and then the ship 031 is extracted.
[0106]
FIG. 26 is a diagram in which the object region 261 of the ship 032 extracted by the method described in the first embodiment is blacked out. For this situation, as in the first embodiment, the schematic shape is set as shown in FIG. 5, and the child blocks are arranged as shown in FIG.
[0107]
And the pixel value of a pixel with a child block is set to X _c , The pixel value of the corresponding pixel of the reduced block is X _p X _c And X _p The error calculation formula is
(1) X _c Or X _p When one of the two is within the area 261, “X _c And X _p Error ”=“ 0 ”
(2) If the above (1) is not met, _c And X _p Error ”=“ X _c And X _p Absolute value of difference in pixel value from
Define as follows.
[0108]
By defining in this way, the error is not counted in the horizontal line region at the upper left of the image pattern 112 in FIG. 11, the error of the block 093 becomes zero, and this is selected as the parent block. That is, the error calculation formula is such that a parent block 093 similar to the child block 091 in FIG.
[0109]
As described above, according to the present embodiment, when the contours of a plurality of objects are sequentially extracted, the detected objects are masked using information on the contours of the already detected objects, so that the plurality of objects are close to each other. However, it is possible to accurately extract the contour of the object.
[0110]
(Third Embodiment) A third embodiment of the present invention will be described with reference to the drawings. Conventionally, the present embodiment is intended to solve the problem that it is difficult to detect the contour when the object region is thin. Here, the description will focus on the parts different from the first embodiment.
[0111]
In the conventional fractal method, it is difficult to perform object extraction when the object shape is thin, such as around the child block 161 in FIG.
[0112]
There is no completely similar parent block for the child block 161 arranged in the narrow portion. For the child block 161, for example, the block 162 is selected as the parent block.
[0113]
When the child block 161 and the parent block 162 are used to perform replacement conversion from the parent block 162 to the child block 161 for the alpha mask of the approximate shape 051, the width of the object region of this part is reduced to 1/2 each time. If the replacement conversion is repeatedly performed, the object region will eventually disappear.
[0114]
This embodiment can cope with the above-mentioned problem. Hereinafter, the flow of processing in the present embodiment will be described with reference to the flowchart of FIG.
[0115]
(Step 020) The same processing as step 010 of the first embodiment is performed.
[0116]
(Step 021) Child blocks are arranged along the outline of the outline-shaped alpha mask as shown in FIG.
(Step 022) Of the child blocks arranged in step 021, those determined to be arranged in a thin portion of the object are removed. A method for determining whether or not a thin portion will be described later.
[0117]
(Step 023) A parent block similar to the child block not canceled in step 022 is found with reference to the image data. In the present embodiment, similar blocks are obtained while referring to label data as in the first embodiment.
(Step 024) Step 023 is repeated until parent blocks for all child blocks are determined.
[0118]
(Step 025) On the alpha mask, the data of the area corresponding to each child block is replaced with the data of the area of the corresponding parent block reduced to the same number of pixels as the child block.
(Step 026) Step 025 is performed for all child blocks. As a result, all child blocks in the alpha mask are replaced once.
(Step 027) Steps 025 and 026 are recursively repeated a predetermined number of times. Thereby, the contour of the alpha mask asymptotically approaches the contour of the object.
[0119]
In this embodiment, label data is used, but it is not essential. It is not necessary to use label data as in the past.
[0120]
By processing in this way, since the child block is not arranged in the thin object portion, the problem that the portion disappears can be avoided in many cases.
[0121]
The determination in step 022 as to whether or not the object is a thin portion is performed using the pixel value in the alpha mask of the child block to be determined. Hereinafter, a specific processing method will be described with reference to FIGS. 18, 19, 20, 21, and 22.
[0122]
FIG. 18 is an enlarged view of an alpha mask of a child block of 16 × 16 pixels. White pixels represent the object, and black pixels represent the background.
[0123]
The child block alpha mask is scanned one pixel at a time from the upper left pixel to the upper right pixel, and the number of pixel value changes is counted. Further, four sides are scanned from the upper right end pixel to the lower right end pixel, the lower right end pixel to the lower left end pixel, and the lower left end pixel to the upper left end pixel, and all the changes are counted.
[0124]
The number of changes required in this way is always an even number. And when the frequency | count of a change is 4 times or more, it determines with there being a thin part of an object.
[0125]
In the case of FIG. 18, the scanning from the upper left pixel to the upper right pixel and the scanning from the lower right pixel to the lower left pixel each change twice from black to white and from white to black. Since it is a total of 4 times, it is determined to be a thin portion.
[0126]
Since the child block is arranged along the outline of the general shape, when the periphery of the block is scanned in this way, if it is not a thin portion, the number of changes as shown in FIG. 21 is usually two. When the number of changes is two, it can be extracted by the fractal method.
[0127]
When the number of changes is more than 2, it is estimated that the shape is a complicated shape or a thin portion. In the case of complicated shapes, the fractal method often cannot cope.
[0128]
FIG. 19 shows an example in which the number of changes is 6 and FIG. 20 shows an example in which the number of changes is 4. In such a case, since there is no self-similarity, it is considered better to avoid extinction of the object region by canceling the child block and leaving the approximate shape as it is without forcibly applying the fractal method.
[0129]
In the case shown in FIG. 22, the number of changes is 0. In this case, the white object region disappears due to replacement conversion. If this small object is noise and can be eliminated, the child block can be left as it is, but if it is known that this is a small object, the user can specify it with the mouse, etc. Give this information to cancel this child block.
[0130]
In the above example, four sides of the child block are scanned in step 022. Although this method is simple and easy, the disappearance of a thin object portion cannot be completely avoided.
[0131]
For example, even if the child block pattern is FIG. 21 (a), if the parent block is the one shown in FIG. 20 (a), the child block is overwritten by one replacement conversion, and the replacement conversion is performed. The object region disappears by repeating the above.
[0132]
In order to avoid such a situation, a parent block is first obtained, and four sides of the parent block are scanned to count the number of changes in the pixel value of the alpha mask. If the number of changes is greater than 2, the parent block and child block are canceled.
[0133]
A flowchart in this case is shown in FIG. Except for the step 234 for removing the child block and the parent block, the process is the same as in FIG.
[0134]
FIG. 17 is a diagram illustrating an example of the positional relationship between the child block 161 and the parent block 162. So far, the example of scanning the four sides of the child block 161 and the example of scanning the parent block 162 have been described.
[0135]
In addition, the four sides of the block 163 that shares the center with the child block and has the same size as the parent block 162 may be scanned, and similarly, it may be determined whether or not the portion is a thin portion depending on the number of changes. In the case of the block 163, since the determination can be made before the parent block search, the processing amount can be reduced. Further, a result closer to the determination using the parent block 162 than using the child block 161 is obtained.
[0136]
As described above, in this embodiment, the pixels on the four sides of the parent block are scanned, the shape of the object included in the parent block is estimated, and it is determined whether or not to perform the replacement process. It is possible to leave a contour such as a thin part or a part having a complicated shape.
[0137]
(Fourth Embodiment) An object region that disappears by repetition of replacement conversion is an object region having a shape narrower than the width of the parent block. Therefore, in order to avoid the disappearance of the object area, the tolerance of the schematic shape is reduced, but it is considered that the parent block and the child block should be reduced.
[0138]
However, when the child block is originally a small block such as 4 × 4 pixels, if it is made smaller than this, the resolution becomes coarse and an image pattern sufficient to use self-similarity cannot be obtained.
[0139]
Therefore, in this embodiment, disappearance is avoided by enlarging the original image and the schematic shape without changing the size of the parent block, and making the object region thicker than the width of the parent block. If necessary, the extracted alpha mask is reduced to the original size.
[0140]
A flowchart of this embodiment is shown in FIG. The difference from the previous embodiments is that step 271 for enlarging the original image and the schematic shape is performed at the very beginning of the processing. The enlargement of the original image in step 271 is performed using, for example, zero-order interpolation, linear (primary) interpolation, or a higher-order pixel interpolation method. For enlargement of the approximate shape, which is an alpha mask, for example, zero-order interpolation is used. Subsequent steps are the same as those in the previous embodiments, and a description thereof will be omitted.
[0141]
As described above, in this embodiment, the contour is extracted by enlarging the original image. Therefore, the contour of a thin part of an object can be accurately extracted even if the shape is likely to disappear in the conventional contour extraction processing method. It becomes like this.
[0142]
(Fifth Embodiment) This embodiment relates to an outline extraction process suitable for extracting the outline of an object at the edge of an image.
[0143]
In the fractal method, high extraction accuracy may be obtained by padding a certain width outside the original image. For example, as shown in FIG. 24, it is assumed that an object 242 exists in a hatched portion in the original image 241 and a rough shape 243 slightly deviated from the outline of the object 242 is given.
[0144]
In the case of the child block 244 arranged as shown in FIG. 24, in order to bring the schematic shape 243 closer to the contour of the object 242, the parent block is positioned so that the image pattern inside thereof is similar to the child block. It is desirable to arrange in block 245.
[0145]
Since the parent block search is performed within a certain range centered on the child block, it is sufficient that the block 245 exists in the range. However, in the case of FIG. 24, the right side of the block 245 protrudes from the image. Therefore, even if the block 245 is within the search range, error evaluation is not performed and it is excluded from the parent block candidates. Therefore, in the child block 244 portion, a block different from that shown in the figure is selected as the parent block, and the contour cannot be obtained accurately.
[0146]
Therefore, in the present embodiment, the image is expanded in advance to the extent that there is a possibility of becoming a search range, as in a region 253 indicated by a horizontal line in FIG. This is called padding.
[0147]
In the present embodiment, the padding pixel value is extrapolated by the outermost pixel value of the image, and the position and width for padding are determined as follows.
[0148]
When it is known in advance that the child block is arranged on the outline of the approximate shape, a rectangle 251 circumscribing the approximate shape 243 is first obtained prior to object extraction. If the rectangle 251 is obtained, the outer frame 252 of the portion that can be the parent block search range when the child block is arranged around the inside is determined.
[0149]
The inside of the outer frame 252 may calculate an error as a candidate block when searching for a parent block. Therefore, padding is performed for the width of the outer frame 252 beyond the image 241. At this time, padding is not required for the portion where the outer frame 252 does not protrude, such as the upper left or lower right side of the image 241, but if there is no particular limitation on the memory capacity for storing the image data, Padding with a sufficient width may be performed on all the left and right sides. By padding the outside of the image in this way, it is possible to avoid the problem that the contour line near the edge of the image cannot be accurately extracted.
[0150]
FIG. 28 is a flowchart of the object extraction process of this embodiment. At the beginning of object extraction by the so-called fractal method, step 281 is executed to perform padding. Other steps are the same as in the previous embodiments.
[0151]
As described above, in the present embodiment, by providing padding with a predetermined width outside the original image, the contour existing at the edge of the image can be accurately extracted.
[0152]
(Sixth Embodiment) This embodiment is an embodiment relating to the contour extraction process in the case where there is a large deviation between the approximate shape and the correct shape.
[0153]
FIG. 30 is an example of a case where an object contour cannot be correctly extracted by the conventional fractal method. FIG. 30 is a diagram showing a contour 301 of an object, its approximate shape 302, and a child block 303 that is one of the child blocks arranged so that the center is located on the contour of the approximate shape 302.
[0154]
Since the child block 303 does not include the object contour 301, for example, the block 311 shown in FIG. 31 is selected as a similar block to the child block 303 in the parent block search. Thus, when the gap between the approximate shape and the correct shape is large, the self-similar structure of the object cannot be extracted, and as a result, the contour of the object may not be extracted correctly.
[0155]
Therefore, in this embodiment, the parent block is arranged first. This is shown in FIG. 32 and FIG. First, the parent block is arranged so that the center is located on the outline of the schematic shape 302 as in the block 321 of FIG. Then, a 1/2 vertical and horizontal block similar to the parent block is searched.
[0156]
In this case, the object outline 301 is included in the lower left of the block 321. Therefore, when a vertical and horizontal ½ block similar to the block 321 is searched, the block 331 is selected as a child block as shown in FIG.
[0157]
FIG. 29 is a flowchart of the contour extraction process of the present embodiment. Steps 291 to 293 are processing for searching similar child blocks by arranging the parent blocks first. Subsequent replacement processing is the same as that of the previous embodiment.
[0158]
(Step 291) The parent block P is arranged in the contour portion of the alpha mask. Similar to FIG. 7 showing an example of arranging child blocks, the outlines of the parent blocks P are arranged so that outlines of the outline shape pass, and adjacent parent blocks overlap.
(Step 292) A child block C similar to the parent block P is found.
(Step 293) Step 292 is performed on all parent blocks to determine child blocks.
Since the following is a replacement process similar to the above, description thereof will be omitted.
[0159]
As described above, in the present embodiment, the parent block is first arranged on the outline of the approximate shape, and the child block is searched later, so that the allowable error between the approximate shape and the correct shape is higher than when the child block is arranged first. Therefore, it becomes easier to extract the self-similar structure of the object than before. As a result, the object can be easily extracted correctly.
[0160]
(Seventh Embodiment) This embodiment is an embodiment relating to a contour extraction process suitable for obtaining an alpha mask that also represents an intermediate state between an object region and a background region. Hereinafter, this embodiment will be described.
[0161]
FIG. 36 is a flowchart of contour extraction processing according to this embodiment.
(Step 361) As shown in FIG. 34, the enlarged image data 344 and the enlarged alpha mask 343 are generated by enlarging the image data 342 and the substantially shaped alpha mask 341 at a predetermined magnification. In the present embodiment, the enlargement ratio is doubled vertically and horizontally.
(Step 362) The above-described fractal method is performed on the enlarged image data 344 and the enlarged alpha mask 343 to extract an accurate alpha mask. Here, the pixel value of each pixel of the alpha mask is 0 for the background and 255 for the object.
(Step 363) The alpha mask obtained by the fractal method is reduced to the original size to obtain an alpha mask for the image data 342.
[0162]
FIG. 35 is a diagram for explaining the reduction processing of the enlarged alpha mask to the original size in step 363. In the present embodiment, since the original size image is enlarged twice and vertically and the fractal method is performed, the four pixels 353 of the enlarged alpha mask correspond to the pixels 352 of the alpha mask obtained by reduction. In the present embodiment, the average value of the pixel values of the pixel 353 is used as the pixel value of the pixel 352 of the alpha mask obtained by reduction. Accordingly, the pixel value of each pixel of the alpha mask obtained by reduction is an average value of four pixels.
[0163]
In this embodiment, each pixel of the enlarged alpha mask is given a pixel value of 0 for the background and a pixel value of 255 for the object. Therefore, among the four enlarged alpha mask pixels corresponding to each pixel of the alpha mask obtained by reduction, the average value is 0, 1, 2, 3, and 4, respectively. , 0, 64, 128, 192, 255.
[0164]
In FIG. 35, since the pixel 352 includes two object pixels, the pixel value of the pixel 352 is 128. Further, if the values of the corresponding pixels are all the same as in the pixel 354, the corresponding pixels 351 have the same value.
[0165]
By using the pixel value of each pixel of the alpha mask obtained in this way as the pixel mixture ratio when combining the extracted object with another image, the jagged feeling given by the contour line is improved and more natural composition An image can be obtained.
[0166]
The calculation formula of the pixel value of the composite image is as follows: X is the pixel value of the pixel of the image in which the extracted object is shown, A is the value of the alpha mask corresponding to that pixel, and the pixel at the same position by another image to be combined If the pixel value is Y,
“Pixel value of composite image” = (A × X + (255−A) × Y) / 255
It becomes like this. In this way, the pixel value of each pixel of the alpha mask can be regarded as representing transparency with 0 being completely transparent and 255 being completely opaque.
[0167]
In the above-described enlargement process, pixel interpolation is performed on the image data by, for example, first-order interpolation, and on the alpha mask, for example, by zero-order interpolation. For example, when generating an alpha mask for a reduced image called a thumbnail image, it is not necessary to enlarge the original image, and after generating an alpha mask having values of 0 and 255 at the original size, You can reduce it as you did.
Further, if an image corresponding to transparency, such as image compression standards PNG and GIF, is generated and stored using the image data and the generated alpha mask, for example, it becomes a general-purpose object image and can be reused.
[0168]
As described above, in the present embodiment, when generating an alpha mask by reducing the enlarged alpha mask obtained by enlarging the original image and performing contour extraction, the multistage obtained by referring to the plurality of pixel values of the enlarged alpha mask Is used as each pixel value of the alpha mask. Therefore, by using the pixel value of the alpha mask as the transparency when extracting the object, it is possible to obtain a smooth contour line of the object.
[0169]
【The invention's effect】
As described above, according to the present invention, in the fractal method, the contour of an object can be accurately obtained even when the edge of another object contour or the like is near or the object region is thin.
[Brief description of the drawings]
FIG. 1 is a flowchart of contour extraction processing according to a first embodiment of the present invention.
FIG. 2 is a flowchart of contour extraction processing according to a third embodiment of the present invention.
FIG. 3 shows an example of an extraction target image.
FIG. 4 is an example of an alpha mask for the image of FIG.
FIG. 5 shows a schematic shape set.
FIG. 6 is a schematic shape alpha mask.
FIG. 7 shows the arrangement of child blocks.
FIG. 8 shows an example of a child block and its parent block.
FIG. 9 shows an example of another child block and its parent block.
FIG. 10 is an example in which another parent block is selected in FIG. 9;
11 shows how an error is calculated in FIG.
FIG. 12 shows how an error is calculated in FIG.
FIG. 13 shows an example in which an error is calculated according to the present invention.
FIG. 14 shows another example in which an error is calculated according to the present invention.
FIG. 15 shows an example of inputting label data.
FIG. 16 shows an example in which child blocks are arranged in a narrow portion.
FIG. 17 shows an example of a child block arranged in a thin portion and a surrounding parent block.
FIG. 18 is an alpha mask in a child block.
FIG. 19 shows an example of an alpha mask pattern: the number of changes is 6.
FIG. 20 shows an example of an alpha mask pattern: the number of changes is 4.
FIG. 21 shows an example of an alpha mask pattern: the number of changes is 2.
FIG. 22 shows an example of an alpha mask pattern: the number of changes is zero.
FIG. 23 is a flowchart of a modification of the contour extraction process according to the third embodiment of the present invention.
FIG. 24 shows an example in which an object is at the edge of an image.
FIG. 25 shows an example in which padding is provided at the edge of an image for expansion.
FIG. 26 shows an example in which a first object region is extracted and a mask is set.
FIG. 27 is a flowchart of contour extraction processing according to the fourth embodiment of the present invention;
FIG. 28 is a flowchart of contour extraction processing according to the fifth embodiment of the present invention;
FIG. 29 is a flowchart of contour extraction processing according to the sixth embodiment of the present invention.
FIG. 30 shows an example in which child blocks are arranged.
FIG. 31 shows an example of obtaining a parent block.
FIG. 32 shows an example in which parent blocks are arranged.
FIG. 33 shows an example of obtaining a child block.
FIG. 34 is a diagram for explaining the concept of processing for enlarging image data and an alpha mask.
FIG. 35 is a diagram for explaining the concept of alpha mask reduction processing;
FIG. 36 is a flowchart of contour extraction processing according to the seventh embodiment of the present invention;
FIG. 37 shows an example in which expansion and contraction are performed in a schematic shape.
FIG. 38 is a diagram showing a sampling area when creating a color table.
FIG. 39 shows an example of creating a color table locally.
FIG. 40 is a diagram illustrating a schematic configuration of a computer used in each embodiment of the present invention.
[Explanation of symbols]
031 Ship for contour extraction
032 Another ship in the vicinity of ship 031
051 Outline shape of ship 031
081, 091, 161, 244, 303, 331 Child block
082, 093, 121, 162, 245, 311, 321 Parent block
114 Part of the outline of ship 032
115 Part of the outline of ship 031
111, 132 Child block
112, 133 Parent block reduced
113, 122, 134 Difference
131, 135 Part of the border of the label data
151 Label data boundary
241 images
242 objects
243 Schematic shape of object 242
253 Extension area by padding process
301 Part of object contour
302 Part of rough shape

Claims (14)

A first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data;
A second step of setting a plurality of search reference blocks by shifting their positions along the outline of the schematic shape represented by the shape data;
For each of the search reference blocks, the design of the area that is larger than the search reference block and that corresponds to this block on the image data is most similar to the design of the area that corresponds to the search reference block on the image data. A third step of searching for similar blocks,
A process of replacing the shape data of the area corresponding to each search reference block with a reduced version of the shape data of the area corresponding to each similar block, and repeatedly performing a process until a predetermined condition is satisfied,
In the first step,
Also receives label data representing the certainty of the target object at each pixel in the image data,
In the third step,
Referring to the label data, ignore or neglect the error between pixels that are unlikely to contain the target object, calculate the weighted sum of errors of the entire block, and search for similar blocks that minimize the weighted sum Do,
Object image extraction method. The shape data is
Represents a region of the approximate shape of the target object, each pixel corresponding to each pixel of the image data one-to-one,
The label data is
The image data is divided into at least two types of three types of target object determination region, non-target object determination region, and undetermined region, and each pixel has a one-to-one correspondence with each pixel of the image data And
In the third step,
For each of a plurality of candidate blocks that are candidates for the similar block,
Calculating the error of each corresponding point between the data of the region corresponding to the search reference block in the image data and the data obtained by reducing the region corresponding to the candidate block in the image data;
The weighted sum is calculated when the corresponding points are both present in a region that is classified as a target object determination region by the label data, or when both are present in a region that is classified as a non-target object determination region. Ignore or neglect errors about corresponding points,
Determining a candidate block having the smallest weighted sum of errors as the similar block;
The object image extraction method according to claim 1. In the third step,
For each of a plurality of candidate blocks that are candidates for the similar block,
Calculating the error of each corresponding point between the data of the region corresponding to the search reference block in the image data and the data obtained by reducing the region corresponding to the candidate block in the image data;
In the case where both of the corresponding points are present in the region classified into the target object fixed region by the label data, or both are present in the region classified into the non-target object fixed region, Obtain a weighting factor by which the error for the corresponding point is multiplied by 0 or a small value,
Determining a candidate block having the smallest weighted sum of errors as the similar block;
The object image extraction method according to claim 2. A first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data;
A plurality of search reference blocks are set to be shifted from each other along the outline of the schematic shape represented by the shape data, and among these, the pattern of the region corresponding to the search reference block on the shape data is A pattern of a thin portion of the target object is a second step of canceling the setting;
For each of the search reference blocks, a search is made for a similar block that is larger than the search reference block and whose symbol in the region corresponding to that block on the image data is most similar to the symbol in the region corresponding to the search reference block. And the steps
A fourth step of repeatedly performing the process of replacing the data of the area corresponding to each search reference block with the reduced data of the area corresponding to each similar block in the shape data until a predetermined condition is satisfied;
Have
In the second step, the number of changes in the pixel value is obtained by scanning the periphery of the region corresponding to the search reference block on the shape data, and when the number of changes exceeds two, Determining that the pattern of the region corresponding to the search reference block is a pattern of a thin portion of the target object;
Object image extraction method. In the second step,
A plurality of search reference blocks are set by shifting their positions along the outline of the schematic shape represented by the shape data,
For the inspection block having the same center position as each search reference block, examine the pattern of the area corresponding to the inspection block on the shape data,
Cancel the setting of the search reference block when the pattern of the inspection block is the pattern of the thin part,
The object image extraction method according to claim 4 . In the second step, the number of changes in the pixel value is obtained by scanning the periphery of the region corresponding to the inspection block on the shape data. If the number of changes exceeds two, Determining that the pattern is a pattern of a thin portion of the target object;
The object image extraction method according to claim 5 . A first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data;
A second step of setting a plurality of search reference blocks by shifting their positions along the outline of the schematic shape represented by the shape data;
For each search reference block, search for a similar block that is larger than the search reference block and whose pattern on the image data corresponds most to the pattern of the area corresponding to the search reference block, and A third step of canceling the setting of a similar block in which a pattern of a corresponding region on the shape data is a pattern of a thin portion of the target object and a search reference block corresponding thereto;
A fourth step of repeatedly performing the process of replacing the data of the area corresponding to each search reference block with the reduced data of the area corresponding to each similar block in the shape data until a predetermined condition is satisfied;
Have
In the third step, the circumference of the area on the shape data corresponding to the similar block is scanned to determine the number of pixel value changes, and if the number of changes exceeds two, the similarity block Determining that the corresponding pattern of the region on the shape data is a pattern of a thin portion of the target object;
Object image extraction method. A first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data;
A second step of setting a plurality of search reference blocks by shifting their positions along the outline of the schematic shape represented by the shape data;
For each search reference block, a search is made for a similar block that is smaller than the search reference block and whose pattern in the image data corresponding to the block is most similar to the pattern in the area corresponding to the search reference block. And the steps
A fourth step of repeatedly performing the process of replacing the data of the area corresponding to each similar block in the shape data with the reduced data of the area corresponding to each search reference block until a predetermined condition is satisfied;
An object image extraction method comprising: First image data including image data including a target object to be extracted and shape data representing a region of a schematic shape of the target object in the image data are generated to generate enlarged image data and enlarged shape data. Steps,
A second step of setting a plurality of search reference blocks by shifting their positions along the outline of the schematic shape represented by the enlarged shape data;
For each search reference block, a search is made for a similar block that is larger than the search reference block and whose symbol on the enlarged image data is most similar to the symbol of the region corresponding to the search reference block. 3 steps,
A fourth step of repeatedly performing the process of replacing the data of the area corresponding to each similar block with the reduced data of the area corresponding to each search reference block in the enlarged shape data until a predetermined condition is satisfied;
A fifth step of reducing the enlarged shape data obtained in the fourth step to the original size;
An object image extraction method comprising: On the computer,
A first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data;
A second step of setting a plurality of search reference blocks by shifting their positions along the outline of the schematic shape represented by the shape data;
For each of the search reference blocks, a block that is larger than the search reference block and whose pattern in the region corresponding to the block on the image data is most similar to the pattern in the region corresponding to the search reference block A third step of searching;
A step of repeatedly performing the process of replacing the shape data of the area corresponding to each search reference block with the reduced shape data of the area corresponding to each similar block until a predetermined condition is satisfied, and
In the first step,
Also receives label data representing the certainty of the target object at each pixel in the image data,
In the third step,
Referring to the label data, ignore or neglect the error between pixels that are unlikely to contain the target object, calculate the weighted sum of errors of the entire block, and search for similar blocks that minimize the weighted sum Do,
Object image extraction program. The shape data is
Represents the schematic shape of the target object, each pixel of which is one-to-one correspondence with each pixel of the image data,
The label data is
The image data is divided into at least two types of three types of target object determination region, non-target object determination region, and undetermined region, and each pixel has a one-to-one correspondence with each pixel of the image data age,
In the third step,
For each of a plurality of candidate blocks that are candidates for the similar block,
Calculating the error of each corresponding point between the data of the region corresponding to the search reference block in the image data and the data obtained by reducing the region corresponding to the candidate block in the image data;
If the corresponding points are both present in a region that is classified as a target object defined region by the label data, or both are present in a region that is classified as a non-target object defined region Ignore or neglect errors about points,
Determining a candidate block having the smallest weighted sum of errors as the similar block;
The object image extraction program according to claim 10 . On the computer,
A first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data;
A plurality of search reference blocks are set to be shifted from each other along the outline of the schematic shape represented by the shape data, and among these, the pattern of the region corresponding to the search reference block on the shape data is A pattern of a thin portion of the target object is a second step of canceling the setting;
For each of the search reference blocks, a search is made for a similar block that is larger than the search reference block and whose symbol in the region corresponding to that block on the image data is most similar to the symbol in the region corresponding to the search reference block. And the steps
A fourth step of repeatedly performing the process of replacing the data of the area corresponding to each search reference block with the reduced data of the area corresponding to each similar block in the shape data until a predetermined condition is satisfied;
The execution,
In the second step, the number of changes in the pixel value is obtained by scanning the periphery of the region corresponding to the search reference block on the shape data, and when the number of changes exceeds two, Determining that the pattern of the region corresponding to the search reference block is a pattern of a thin portion of the target object;
Object image extraction program. In the second step,
A plurality of search reference blocks are set by shifting their positions along the outline of the schematic shape represented by the shape data,
For the inspection block having the same center position as each search reference block, examine the pattern of the area corresponding to the inspection block on the shape data,
Cancel the setting of the search reference block when the pattern of the inspection block is the pattern of the thin part,
The object image extraction program according to claim 12 . On the computer,
A first step of receiving image data including a target object to be extracted, and shape data representing a region of a schematic shape of the target object in the image data;
A second step of setting a plurality of search reference blocks by shifting their positions along the outline of the schematic shape represented by the shape data;
For each search reference block, search for a similar block that is larger than the search reference block and whose pattern on the image data corresponds most to the pattern of the area corresponding to the search reference block, and A third step of canceling the setting of a similar block in which a pattern of a corresponding region on the shape data is a pattern of a thin portion of the target object and a search reference block corresponding thereto;
A fourth step of repeatedly performing the process of replacing the data of the area corresponding to each search reference block with the reduced data of the area corresponding to each similar block in the shape data until a predetermined condition is satisfied;
The execution,
In the second step, the number of changes in the pixel value is obtained by scanning the periphery of the region corresponding to the search reference block on the shape data, and when the number of changes exceeds two, Determining that the pattern of the region corresponding to the search reference block is a pattern of a thin portion of the target object;
Object image extraction program.

Images