WO2011128933A1

WO2011128933A1 - Image vectorization device, image vectorization method, and image vectorization program

Info

Publication number: WO2011128933A1
Application number: PCT/JP2010/002671
Authority: WO
Inventors: 田口進也
Original assignee: 三菱電機株式会社
Priority date: 2010-04-13
Filing date: 2010-04-13
Publication date: 2011-10-20

Abstract

An image vectorization training unit (3) converts a sample raster image in a raster image database (2) to a vector image in accordance with a given method, and stores the vector image paired with an image feature value into a pattern database for raster to vector conversion (4). An image vectorization execution unit (5) divides an inputted raster image into portions, selects from the pattern database for raster to vector conversion (4) image vectors which are stored paired with the image feature values conforming to image feature values of the respective portions, and combines the image vectors.

Description

Image vectorization apparatus, image vectorization method, and image vectorization program

The present invention relates to, for example, an image vectorization apparatus, an image vectorization method, and an image vectorization program for converting a raster image into a vector image format.

Hereinafter, the “vector image” of the present invention refers to an image in which an image is represented by dots, lines, and surfaces, and color information is represented by a parametric equation. Here, “point” refers to a two-dimensional coordinate value, a three-dimensional coordinate value, or a more generally expressed N-dimensional coordinate value. A “line” refers to a straight line or a curve connecting two points. Further, the “surface” refers to a region surrounded by a plurality of lines. A “raster image” refers to an image represented by a set of colored points. Also, “image vectorization” refers to converting a raster image into a vector image.

Conventionally, image vectorization methods have been proposed for the purpose of facilitating image editing and preventing the shape in the image from being blurred or distorted even when the image is enlarged or reduced (for example, non-patented). Reference 1). FIG. 13 shows an example of image vectorization by the method of Non-Patent Document 1. In each graph of FIGS. 13A and 13B, the horizontal axis represents image coordinates, and the vertical axis represents pixel values (color information). For simplification of description, the two-dimensional image coordinates are expressed by one-dimensional image coordinates. As shown in FIG. 13A, when there is no noise in the pixel value of the raster image, image vectorization is realized by approximating each pixel value with a smooth bicubic surface L. A portion surrounded by a broken line in FIG. 13A is a portion where the color changes greatly, and corresponds to, for example, an edge (contour) of an object.

However, in the image vectorization method described in Non-Patent Document 1, image vectorization is performed so as to faithfully reproduce the input raster image. Therefore, even noise existing in the raster image is faithfully reproduced. There was a problem of end. For this reason, for example, as shown in FIG. 13B, when noise is added to the raster image, even the pixel value Z of the noise is tried to be faithfully reproduced. The approximate curved surface L ′ has been reproduced.

Further, in the image vectorization method described in Non-Patent Document 1, it has been a premise for image vectorization that the edge of an object is clearly imaged. Therefore, even in the case of a raster image in which the edge of an object is blurred, there is no processing for emphasizing the edge and converting it to an image vector.

The present invention has been made to solve the above-described problems. An image vectorization apparatus and an image vector that automatically perform image vectorization on a raster image including noise without being affected by noise. An object is to obtain a conversion method and an image vectorization program.

An image vectorization apparatus according to the present invention includes an image vectorization learning unit that converts a sample raster image into a vector image according to a predetermined method, and an image vectorization pattern that holds a pair of the sample raster image and the vector image A database and an image vectorization execution unit that converts an input raster image into a vector image using a sample raster image and a pair of the vector images held in the image vectorization pattern database are provided.

Further, the image vectorization method of the present invention uses an image vectorization learning step for converting a sample raster image into a vector image according to a predetermined method, and a sample raster image and a vector image pair. And an image vectorization execution step for converting the raster image into a vector image.

The image vectorization program according to the present invention uses an image vectorization learning procedure for converting a sample raster image into a vector image according to a predetermined method, and uses a sample raster image and a vector image pair. An image vectorization execution procedure for converting the raster image into a vector image is executed by a computer.

According to the present invention, by learning the image vectorization method in advance using the raster image for the sample, the image vectorization is automatically performed on the raster image including the noise without being affected by the noise. Can be implemented.

It is a block diagram which shows the structure of the image vectorization apparatus which concerns on Embodiment 1 of this invention. 6 is a flowchart illustrating an operation of an image vectorization learning unit of the image vectorization apparatus according to Embodiment 1; FIG. 3A shows a conversion process of the image vectorization learning unit, FIG. 3A shows a raster image before conversion, and FIG. 3B shows a vector image after conversion. It is a figure which shows an example of the local coordinate system defined to the image. It is a figure which shows the example which image-converts the raster image to which noise was added by the image vectorization learning part. It is a block diagram which shows the structure of an image vectorization pattern database. 4 is a flowchart illustrating an operation of an image vectorization execution unit of the image vectorization apparatus according to Embodiment 1; It is a figure which shows the example of calculation of the image feature-value by an image vectorization execution part. It is a figure which shows the candidate vector image selected by the image vectorization execution part, and its candidate probability. It is a figure which shows the selection method of the vector image by an image vectorization execution part. It is a figure which shows the example of the calculation method of the consistency of the boundary of an attention partial area | region and a peripheral partial area | region by the image vectorization execution part. It is a figure which shows the example of a deformation | transformation process of the candidate vector image by an image vectorization execution part. FIGS. 12A and 12B illustrate an image vectorization method according to Non-Patent Document 1. FIG. 12A shows a case of a raster image without noise, and FIG. 12B shows a case of a raster image containing noise.

Hereinafter, in order to explain the present invention in more detail, modes for carrying out the present invention will be described with reference to the accompanying drawings.
Embodiment 1 FIG.
An image vectorization apparatus 1 shown in FIG. 1 includes a raster image database 2 that stores a plurality of raster images as samples for learning a strategy for image vectorization, and an image vectorization that converts sample raster images into vector images. Learning unit 3, image vectorization pattern database 4 for storing a pair of sample raster image and vector image, and image vectorization execution for converting a raster image input using image vectorization pattern database 4 into a vector image It consists of part 5.

In the example of FIG. 1, each of the raster image database 2, the image vectorization learning unit 3, the image vectorization pattern database 4, and the image vectorization execution unit 5, which are components of the image vectorization device 1, , A semiconductor integrated circuit board on which an MPU (Micro Processing Unit) is mounted is assumed. However, the image vectorization apparatus 1 may be configured by a computer. In this case, the processing contents of the raster image database 2, the image vectorization learning unit 3, the image vectorization pattern database 4, and the image vectorization execution unit 5 are described. The image vectorization program stored in the memory is stored in the memory of the computer, and the CPU (Central Processing Unit) of the computer executes the image vectorization program stored in the memory.

Next, the operation of the image vectorization apparatus 1 will be described. First, the first half process of the image vectorization apparatus 1 will be described with reference to the flowchart of FIG.
In step ST1, the image vectorization learning unit 3 acquires a plurality of raster images stored in the raster image database 2, and converts the acquired raster images into vector images.

FIG. 3 is a diagram for explaining the conversion process of the image vectorization learning unit 3. FIG. 3A shows the raster image I before conversion, and FIG. 3B shows the image vector V after conversion. The image vectorization learning unit 3 acquires a raster image I having a size of N × M pixels from the raster image database 2 and converts it into a vector image V using an existing image processing function or the like according to an operator's instruction. By doing so, the image vectorization strategy is learned in advance.
For example, since the raster image I in FIG. 3 is blurred, there is ambiguity in the edge position. Therefore, for example, at the discretion of the operator, the raster image I is converted into a vector image V composed of a plane A1 composed of points P1, P2, P3, and P4 and a plane A2 composed of points P2, P5, P3, and P6. . By this conversion processing, information that the edge position of the raster image I is a curve composed of the points P2 and P3 is held in the vector image V, and the level of edge determination can be learned in advance.

Here, an example of a method for expressing the color information and luminance information of the surface A1 shown in FIG. 3 will be described. FIG. 4 is a diagram showing an example of the local coordinate system defined in the image, and shows the local coordinate system (U, V) of the plane A1 with the point P1 as the origin P0 (0, 0). In the figure, u and v are real parameters taking values between 0 and 1, the coordinates of point P2 are (u = 1, v = 0), the coordinates of point P3 are (u = 1, v = 1), and point P4. Is defined as (u = 0, v = 1). At this time, information such as color and luminance at an arbitrary point P (u, v) in the plane A1 can be expressed by an arbitrary parametric function F (u, v).

Here, as the function F (u, v), for example, “Ferguson patch” described in Non-Patent Document 1 described above may be used, or the following expression (1) that can express a steep edge is used. It may be expressed by a sigmoid function.
F (u, v) = 1 / {1 + exp (a * u + b * v + c)} (1)
Here, a, b, and c are constants and may be arbitrarily specified by the user.

Further, the image vectorization learning unit 3 may generate a vector image by removing noise from the input raster image based on an operator's judgment or the like. Thereby, the noise level can be learned in advance.
FIG. 5 shows an example in which a raster image to which noise is added is converted into an image vector. In FIG. 5, the horizontal axis represents image coordinates, and the vertical axis represents pixel values (luminance information). For simplification of description, the two-dimensional image coordinates are expressed by one-dimensional image coordinates. When there is a luminance value Z to which noise is added in the image, the operator or the like determines that the luminance value Z is noise and ignores it, and the image vectorization learning unit 3 determines that the luminance is determined to be noise. An approximate curved surface L that smoothly connects luminance values other than the value Z is calculated.

In subsequent step ST <b> 2, the image vectorization learning unit 3 calculates an image feature amount from the sample raster image acquired from the raster image database 2.
The image feature amount is arbitrary, such as a luminance value included in the raster image, a color histogram, an edge histogram, a color-edge correlation function, or the like. Further, local feature amounts of raster images such as HOG (Histogram of Oriented Gradients) and SIFT (Scale Invariant Feature Transform) may be calculated.

In the subsequent step ST3, the image vectorization learning unit 3 records in the image vectorization pattern database 4 a pair of the image feature amount extracted from the sample raster image and the vector image obtained by converting the raster image into an image vector. Let
FIG. 6 is a block diagram showing the configuration of the image vectorization pattern database 4. The image vectorization learning unit 3 obtains, for example, a raster image Ij (j = 1,..., N) from the raster image database 2, calculates an image feature amount Fj for the raster image Ij, and the vector image Vj is generated, and the pair of the image feature amount Fj and the vector image Vj is stored in the image vectorization pattern database 4 as an image vectorization pattern.

In the above steps ST1 to ST3, the image vectorization learning unit 3 converts the sample raster image of the raster image database 2 into an image vector and stores it in the image vectorization pattern database 4, thereby including noise and edge determination. Learn information on image vectorization strategies.

Next, the latter half process of the image vectorization apparatus 1 is demonstrated using the flowchart of FIG.
First, in step ST11, the image vectorization execution unit 5 acquires a raster image as an input. Here, the raster image to be acquired may have an arbitrary size.

In subsequent step ST12, the image vectorization execution unit 5 divides the acquired raster image into partial regions. FIG. 8A shows an example of a raster image I acquired by the image vectorization execution unit 5, and FIG. 8B shows an example of partial area division. In the example of FIG. 8, the image vectorization execution unit 5 equally divides the input raster image I into fine rectangular areas. Note that the size and shape of the partial area are arbitrary, and may be any size as long as the image feature amount can be acquired from the partial area.

In subsequent step ST13, the image vectorization execution unit 5 calculates each image feature amount from each partial region of the raster image. For example, as shown in FIG. 8C, the image vectorization execution unit 5 calculates an image feature amount Ki from the i-th partial region Ri. Here, as the image feature amount Ki, the same information as the sample image feature amount Fj calculated in step ST2 described above is calculated.
In step ST13, the image vectorization execution unit 5 repeats the same processing for all the partial areas of the raster image I.

In subsequent step ST14, the image vectorization execution unit 5 searches the image vectorization pattern database 4 using the image feature amount of each partial region, and uses a plurality of sample vector images suitable for each partial region as candidate vector images. Extraction is performed, and the candidate probability of each candidate vector image is calculated.

Here, the process of step ST14 is demonstrated using FIG.8 and FIG.9.
The image vectorization execution unit 5 uses the image feature amount Ki calculated from the i-th partial region Ri shown in FIG. 8 to store N image feature amounts Fj (j = 1) stored in the image vectorization pattern database 4. ,..., N) and the image feature amount Ki, the distance d (i, j) is calculated, and a pair is formed with the image feature amount Fj such that the distance d (i, j) is smaller than a certain value. The vector image Vj is acquired from the image vectorization pattern database 4 and is set as a candidate vector image of the partial region Ri. Further, the image vectorization execution unit 5 calculates the candidate probability p (i, j) of each candidate vector image from the distance d (i, j).

As a method of calculating the distance d (i, j) between the image feature values Ki and Fj, for example, when Ki and Fj are given as a one-dimensional array, the Euclidean distance between Ki and Fj, and the vector inner product of Ki and Fj There is a method for calculating a normalized mutual function of Ki and Fj. Alternatively, when the image feature values Ki and Fj are color histograms, the distance d (i, j) may be calculated by the “Bhatterary distance”.

Further, as a method for calculating the candidate probability p (i, j), for example, there is a method for calculating as shown in the following expression (2) using the distance d (i, j).
p (i, j) = a ′ × exp {−b ′ × d (i, j) × d (i, j)} (2)
Here, a ′ is a constant for normalization, and b ′ is a positive constant determined by the user. The larger the candidate probability p (i, j), the more suitable the image vector Vj is for the partial region Ri. The image vectorization execution unit 5 obtains three candidate vector images V1, V2, and V3 shown in FIG. 9 for the partial region Ri shown in FIG.

In subsequent step ST15, the image vectorization execution unit 5 uses the plurality of candidate vector images and each candidate probability to transform the candidate vector image obtained in step ST14 while maintaining consistency with the peripheral partial region. A vector image of each partial area divided in step ST12 is obtained.

Here, the process of step ST15 will be described with reference to FIGS.
First, the image vectorization execution unit 5 selects one candidate vector image that maximizes the candidate probability of the candidate vector image in each partial region, and assigns it as the initial value of the image vector of each partial region. . In the example of FIG. 10A, it is assumed that a vector image has already been obtained in the partial area Sk (k = 1, 2), which is a peripheral partial area of the partial area Ri of interest. Further, it is assumed that three candidate vector images Vj (j = 1, 2, 3) shown in FIG. 9 are obtained in the target partial area Ri of FIG.

In this case, the image vectorization execution unit 5 selects a vector image most suitable for the target partial region Ri while maintaining consistency with the peripheral partial region Sk (k = 1, 2) as follows, for example. Choose from.

First, an example of a method for calculating consistency between the candidate vector image Vj and the peripheral partial region Sk will be described with reference to FIG.
As shown in FIG. 11, the upper boundary region of the candidate vector image Vj of the target partial region Ri is Y1, the left boundary region is Y2, the lower boundary region of the peripheral partial region S1 is W1, and the right side of the peripheral partial region S2 Let W2 be the boundary region. At this time, the image vectorization execution unit 5 calculates a boundary error f (j) between the candidate vector image Vj and the peripheral partial region Sk as in the following expression (3).
f (j) = (pixel value of W1−pixel value of Y1) × (pixel value of W1−pixel value of Y1)
+ (Pixel value of W2−pixel value of Y2) × (pixel value of W2−pixel value of Y2)
(3)
Here, for example, the pixel value of W1 indicates a pixel value generated in the boundary region W1 of the vector image corresponding to the position of the peripheral partial region S1 (that is, the image feature amount calculated in step ST2). The same applies to W2, Y1, and Y2.

In the above equation (3), when the boundary error f (j) becomes a small value, the difference between the pixel values of the candidate vector image Vj and the candidate vector image set as the initial value in the peripheral partial region Sk is small. Show. Accordingly, it can be said that the continuity between the candidate vector image Vj of the target partial region Ri and the candidate vector images of the peripheral partial region Sk is large and consistent. On the other hand, when the boundary error f (j) has a large value, there is a high possibility that the candidate vector image Vj of the target partial region Ri and the candidate vector images of the peripheral partial region Sk are discontinuous, and the consistency is secured. It can be said that it is not.

Subsequently, after obtaining the boundary error f (j) for all candidate vector images Vj in the target partial region Ri, the image vectorization execution unit 5 then calculates the cost function E (j | i defined in the following equation (4). ) Is maximized, and the candidate vector image Vj of the variable j is adopted as the vector image of the target partial region Ri.
E (j | i) = p (i, j) × q (j) (4)
q (j) = a ″ × exp {−b ″ × f (j)} (5)
Here, p (i, j) is the candidate probability of each candidate vector image obtained by the image vectorization execution unit 5 in step ST14 described above. Further, q (j) is a probability representing the continuity of the boundary, and is defined as in the above equation (5). a ″ is a normalization constant, and b ″ is a positive constant set by the user. The cost function E (j | i) is a function of the variable j. In this example, j is one of 1, 2, and 3 shown in FIG.

There are the following advantages in maximizing the cost function E (j | i) and selecting the vector image of the target partial region Ri.
If the candidate vector image most suitable for the target partial area Ri is simply to be determined, a candidate image vector Vj that maximizes the candidate probability p (i, j) of the candidate vector image of the target partial area Ri is determined. That's fine. However, in this method, the boundary between the target partial region Ri and the peripheral partial region Sk becomes discontinuous, and the probability q (j) expressing the consistency with the peripheral partial region Sk takes a small value, resulting in a cost. The function E (j | i) takes a small value.
That is, by maximizing the cost function E (j | i), the degree of matching of the candidate vector image Vj (represented by the candidate probability p (i, j)) in the target partial region Ri is considered, and the peripheral partial region A vector image corresponding to the target partial region Ri can be obtained while considering consistency with Sk (represented by probability q (j)). Therefore, there is an advantage that the raster image can be converted into an image vector so that the boundary areas of the partial areas are continuous.

Further, the image vectorization execution unit 5 selects a candidate vector image that maximizes the cost function E (j | i), and then corrects the vector image to make the consistency with the peripheral partial region Sk more accurate. It may be. FIG. 12 shows an example of a vector image correction method. As shown in FIG. 12A, when there is a slight difference in the boundary region between the line L1 of the target partial region Ri and the line L2 of the peripheral partial region Sk, the image vectorization execution unit 5 As shown in (b), the line L1 of the target partial region Ri may be deformed to achieve consistency with the peripheral partial region Sk.

In step ST15, the image vectorization execution unit 5 repeats the same processing for all the partial areas Ri (j = 1,..., N) of the raster image I, and obtains an appropriate vector image for each partial area Ri. Ask.
In order to sequentially perform the maximization of the cost function E (j | i) for all the partial areas Ri, the image vectorization execution unit 5 uses an optimization method based on iteration such as belief propagation, for example. The vector image of each partial region Ri may be obtained by maximizing the cost function E (j | i) uniformly with respect to Ri.

In subsequent step ST16, the image vectorization execution unit 5 combines the selected vector images with respect to the partial regions Ri of the raster image I acquired in step ST11, and outputs a vector image V.

As described above, according to the first embodiment, the image vectorization apparatus 1 converts a raster image for a sample into a vector image according to a raster image database 2 that holds a raster image for a sample and a predetermined method, and An image vectorization learning unit 3 that calculates an image feature amount, an image vectorization pattern database 4 that holds a pair of the image feature amount and a vector image, and an image feature amount that matches the image feature amount of the input raster image The image vectorization pattern database 4 is searched, and an image vectorization execution unit 5 that selects a vector image paired with the searched image feature amount as a vector image of the input raster image is provided. Therefore, by creating the image vectorization pattern database 4 in advance, information on how much noise is removed, information on which part of the image is considered as an edge, etc. The raster image can be converted into an image vector without being affected by noise.

Note that the image vectorization apparatus 1 can perform the desired image vectorization by changing the information stored in the image vectorization pattern database 4. For example, the image vectorization learning unit 3 intentionally performs an illustration-like image vectorization in which the number of colors of the image is reduced, and stores the pattern in the image vectorization pattern database 4, whereby the image vectorization execution unit 5 can automatically generate an illustration-like vector image. In addition, by storing a photorealistic image vectorization pattern such as an actual photograph in the image vectorization pattern database 4, a vector image such as an actual photograph can be automatically generated.

Further, according to the first embodiment, the image vectorization execution unit 5 calculates a probability indicating the continuity of the boundary between candidate vector images of a partial region of interest and its surrounding partial regions, and pays attention to this probability. The cost function is calculated by multiplying the candidate probabilities of the candidate vector images in the partial area, and the candidate vector image that maximizes the cost function is selected as the optimum vector image for the partial area to be noticed. For this reason, it becomes possible to consider the matching degree of the candidate vector image in the partial area of interest and the consistency with the peripheral partial areas, and the continuity of the boundary areas of the partial areas can be improved.

As described above, the image vectorization apparatus according to the present invention learns the image vectorization method in advance, so that the image vectorization for image vectorization of an image including noise or an image with blurred edges is performed. Suitable for use in equipment.

Claims

An image vectorization learning unit that converts a raster image for a sample into a vector image according to a predetermined method;
An image vectorization pattern database holding a pair of raster images for the sample and vector images thereof;
An image vectorization apparatus comprising: a raster image for a sample held in the image vectorization pattern database; and an image vectorization execution unit that converts an input raster image into a vector image using a pair of the vector images.
The image vectorization learning unit calculates an image feature amount from the raster image for the sample,
The image vectorization apparatus according to claim 1, wherein the image vectorization pattern database holds an image feature amount of the raster image and a pair of the vector image.
The image vectorization execution unit divides the input raster image into partial areas, calculates image feature amounts, and compares them with the image feature amounts of the sample raster images held in the image vectorization pattern database. 3. The image according to claim 2, wherein a vector image paired with the image feature quantity is selected as a candidate vector image of the partial area, and a candidate probability indicating a degree of fitness of the selected candidate vector image is calculated. Vectorization device.
The image vectorization execution unit uses the candidate vector image for each partial region of the input raster image and its candidate probability, while taking the consistency of the candidate vector image in the partial region of interest and the peripheral partial region, 4. The image vectorization apparatus according to claim 3, wherein an optimal vector image is selected from candidate vector images for the target partial region.
The image vectorization execution unit is configured to continuously match the boundary between candidate vector images of the target partial region and the peripheral partial region in order to ensure the consistency of the candidate vector image between the target partial region and the peripheral partial region. Calculating the probability indicating the sex, multiplying the probability by the candidate probability of the candidate vector image of the target partial region of interest, calculating a cost function, and determining the candidate vector image that maximizes the cost function as the target partial region of interest The image vectorization apparatus according to claim 4, wherein the image vectorization apparatus is selected as an optimal vector image for.
An image vectorization learning step for converting a raster image for a sample into a vector image according to a predetermined method;
An image vectorization method comprising: an image vectorization execution step of converting an input raster image into a vector image using a pair of the raster image for the sample and the vector image thereof.
An image vectorization learning procedure for converting a raster image for a sample into a vector image according to a predetermined method;
An image vectorization program for causing a computer to execute an image vectorization execution procedure for converting an input raster image into a vector image using a pair of the sample raster image and its vector image.