KR20140047331A - Object segmentation using block clustering based on automatic initial region of interest estimation - Google Patents

Abstract

The present invention relates to an automatic object segmentation method for automatically segmenting an object from a background using a block clustering algorithm. The automatic object segmentation method is capable of improving user′s convenience and efficiency by automation to supplement the weakness of providing information on the object by a user in the existing GrabCut implementation. The automatic object segmentation method of the present invention comprises an object interest region estimation step of estimating an interest region of an object for image information including segmentation information of a background and the object. In the object interest region estimation step, cluster dispersion information of an image is analyzed to distinguish the cluster according to the size of a dispersion region, and the image is segmented into the predetermined number of blocks and is determined as an object cluster and a background cluster depending on the object area and the background area within an individual block, and a color mean value of the block.

Description

TECHNICAL FIELD [0001] The present invention relates to an automatic object segmentation method and an automatic object segmentation method using an object region based on block clustering and an automatic object segmentation system using block clustering.

The present invention divides an object from a background automatically by using a block clustering algorithm. By the present invention, the convenience and efficiency of a user can be improved by automating the disadvantage of providing information about an object by a user in a conventional GrabCut implementation I decided to plan. Therefore, according to the present invention, an object having a very complicated background and an object having an unspecified shape is automatically divided into image information used in various fields such as "character "," It is a technology that can be applied and applied effectively.

Recently, high-performance mobile phones (smart phones, smart pads, or netbooks) have appeared and people can acquire high-quality image information through mobile phones. And there are many applications that can apply them. Research to acquire and recognize image information of an undefined object has been actively conducted for the past decade. Good feature information is needed to recognize something. And good feature information comes from fully subdivided objects from the background.

In order to extract an object completely from the background, the user has previously been designated to designate an area of interest corresponding to the object in advance. If the user designates the region of interest corresponding to the object in advance, the user can designate the desired region, but there is a limit to the user designation when processing a large number of images. In particular, when editing a composite image (moving image, etc.) composed of a plurality of images, it is difficult for the user to specify each one individually.

In order to solve the above problems, the present invention has a main object of automatically estimating and automating a region of interest of an object according to the present invention, in which a user provides advance information of an object in advance, which is a disadvantage of the GrabCut algorithm .

Another object of the present invention is to cleanly extract information of an object by dividing information of an object candidate automatically before applying an algorithm for separating objects in an image processing field.

Still another object of the present invention is to automate the division of object information by automatically dividing the object in the image by applying the present invention to the estimation process of the GrabCut algorithm.

According to an aspect of the present invention, there is provided a method for analyzing object candidate information and background information of an image, An object region estimation step of estimating a region of interest of an object with respect to image information including background and object division information; And a GrabCut processing step of dividing the object by the GrabCut algorithm using the information of the ROI of the object analyzed by the object ROI estimation step.

According to another aspect of the present invention, there is provided a method for estimating an ROI of an object with respect to image information including a background and an object division information, Wherein the cluster is divided into a predetermined number of blocks and the object cluster and the background cluster are discriminated according to the object area and the background area in the individual block and the color average value of the block. ≪ / RTI >

In a preferred embodiment of the present invention, the object region of interest estimating step includes: a cluster labeling step of performing a labeling process on the background information, a neutral cluster, and an object cluster for the image information, .

In a preferred embodiment of the present invention, the cluster labeling process includes: analyzing the variance information of each cluster to label the background of the image, the neutral, the analyzed background cluster, the neutral cluster, and the object cluster Information analysis phase; And an outer information analyzing step of analyzing the outer area information of the image.

Also, in a preferred embodiment of the present invention, the analyzing step of analyzing the cluster dispersion information comprises the steps of: discriminating a cluster having a large dispersion area from a background cluster, a neutral cluster, and an object cluster to a background cluster, .

According to a preferred embodiment of the present invention, the object region of interest estimation includes: a block clustering process of dividing an image including information of an object region into a predetermined number of blocks and estimating an ROI of the object in the block- The method comprising the steps of:

In a preferred embodiment of the present invention, the block clustering processing step includes: an object area value in a block corresponding to an object cluster in an individual block; an area per block for calculating a background area value in a block corresponding to a background cluster in an individual block; Calculating step; A block color information calculating step of calculating a distance between the average color value of each block and the average color value of the object cluster, and a distance between the average color value of each block and the average color value of the background cluster; And an individual block labeling step of labeling the individual blocks as object clusters or background clusters by comparing object areas, background areas, and color values in the individual blocks calculated by the block area calculating step and the block color information calculating step .

In addition, the present invention includes an object information analyzer for analyzing object information and background information of an image; An object region of interest estimating an area of interest of an object with respect to image information including background and object division information; And a GrabCut processing unit for dividing the object by the GrabCut algorithm using information of the ROI of the object analyzed by the ROI estimating unit, wherein the ROI estimating unit analyzes the cluster dispersion information of the image, Wherein the cluster is divided into a predetermined number of blocks and the object cluster and the background cluster are discriminated according to the object area and the background area in the individual block and the color average value of the block. System.

In an exemplary embodiment of the present invention, the object region of interest estimation unit may include: a community labeling processing unit for performing a labeling process and an area estimation of an object on a background cluster, a neutral cluster, and an object cluster with respect to image information; And a block clustering processor for dividing the image including the information of the object area into a predetermined number of blocks and estimating the ROI of the object in the block-divided image.

In a preferred embodiment of the present invention, the cluster labeling processing unit includes: a cluster distribution processing unit for analyzing distribution information of each cluster to label the background of the image, the neutral, the background cluster analyzed, the neutral cluster, Analysis section; An outer information analyzing unit for analyzing the outer area information of the image; And a neutral community labeling unit for labeling the neutral community with a background community or an object community.

Furthermore, in a preferred embodiment of the present invention, the block clustering processing unit may include: an object area value in a block corresponding to an object cluster in an individual block; an area calculation unit for calculating a background area value in a block corresponding to the background cluster in the individual block; part; A block color information calculation unit for calculating a distance between a color average value of each block and an average color value of the object cluster and a color value average of each individual block and an average color value of the background group; And an individual block labeling unit for labeling individual blocks by an object cluster or a background cluster by comparing the object area, the background area, and the color values in the individual blocks calculated by the block area calculation unit and the block color information calculation unit .

The excellent effect of the present invention as described above is to automate the process of providing advance information by the user, which is the biggest disadvantage of the GrabCut algorithm, which is useful in the image processing field.

Another advantage of the present invention is that the information of the region of interest of the object is automatically calculated from the input image to perform the image processing. Even if the user manually designates the manual processing, The effect is excellent, with little difference.

The characteristic feature of the present invention for automatically estimating a region of interest of an object in order to divide object information of the input image is applied not only to the embodiment of the GrabCut algorithm, which is a preferred application example, but also to various image processing and image information processing algorithms , The processing efficiency of the object information is increased in the image information.

FIG. 1 is a flowchart illustrating a process of processing an automatic object segmentation method according to the present invention.
FIG. 2 is a diagram illustrating an example of processing results of an object information analysis step to which a GMM processing step is applied among automatic object segmentation methods according to the present invention.
FIG. 3 is an exemplary view illustrating a state in which an image is classified into an object cluster, a background cluster, and a neutral cluster by the cluster labeling processing step in the object ROI estimation step in the automatic object division method according to the present invention.
FIG. 4 is an exemplary view illustrating a state in which an image is divided into a plurality of blocks and classified into an object cluster and a background cluster by the block clustering processing step in the object ROI estimation step in the automatic object division method according to the present invention.
FIG. 5 is a diagram illustrating the results of the object information analysis step and the GrabCut processing step after the object ROI estimation step in the automatic object division method according to the present invention.
FIG. 5A is an input image, FIG. 5B is an image according to the scatter information and outer information analysis of the cluster information, FIG. 5C is an example of block clustering, and FIG. (e) is an example of setting an area of interest of an object, and (f) is a GrabCut result image.
6 is a block diagram of an automatic object segmentation system according to the present invention.
7 is a flowchart illustrating an automatic object segmentation method according to the present invention.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

That is, the automatic object segmentation method and the automatic object segmentation system 10 using GrabCut-based Block Clustering according to the present invention can be applied to image information including background and object segmentation information as shown in FIGS. 1 to 7, And an object region of interest estimation step S300 for estimating a region of interest.

That is, an object interest region estimation step (S300) is performed to discriminate an image of an object to be analyzed with respect to a whole image to be input as shown in the lower image of FIG. For this purpose, the object corresponding to the object in the image information is estimated as the region of interest, thereby easily identifying and processing the object in the image processing process after the object region of interest estimation (S300).

In step S300, the object region of interest is analyzed to classify the clusters according to the size of the distributed region, the image is divided into a predetermined number of blocks, the object area and the background area in the individual block, The object cluster and the background cluster are determined according to the color average value of the object cluster and the background cluster.

The automatic object segmentation method according to the present invention is based on GrabCut. An example of the entire process of the image according to the GrabCut process can be applied as an embodiment in which the object interest region estimation step (S300) is included and can be used as follows. Of course, the use of the K-means algorithm, which will be described later, and the GMM process are used to perform the pre-processing for the object region estimation according to the present invention, and the object region of interest The execution of the GrabCut algorithm using the estimated data is a preferred embodiment of the input image processing in the present invention. The application of the automatic object segmentation technique according to the present invention may be implemented by applying the technique of the present invention to various image information processing including K-means algorithm, GMM processing, GrabCut algorithm and the like.

First, an object information analysis step (S200) for analyzing object information and background information of an image is performed according to a preferred embodiment of the present invention.

The object information analysis step S200 may further include a K-means clustering step S210 for performing a clustering algorithm for a target image by the K-means algorithm, a GMM according to a mixture weight, And a GMM processing step S220 of dividing the image into a "background", a "neutral", and an "object" by performing an algorithm.

The object information analysis unit 20 analyzes the object information and the background information of the image. The object information analysis step S200, the K-means clustering step S210, the GMM A processing step S220, and the like will be performed.

Next, an object region of interest estimation step (S300) for estimating an ROI of the object with respect to the image information including the background and object division information is performed, and a detailed description of the object region estimation step (S300) .

Also, a GrabCut processing step (S400) for dividing an object by the GrabCut algorithm using the information of the ROI of the object analyzed by the object ROI (S300) is performed.

A GrabCut processing unit 40 is provided for dividing an object by the GrabCut algorithm using the information of the ROI of the object analyzed by the object ROI estimating unit 30. The GrabCut processing unit 40 performs a GrabCut processing The processing step S400 will be performed.

In addition, an object interest region estimating unit 30 for estimating a region of interest of the object with respect to image information including background and object division information by performing the object region of interest estimation step S300 will be provided.

That is, the object interest region estimating unit 30 analyzes the cluster dispersion information of the image to discriminate the clusters according to the size of the dispersion region, divides the image into a predetermined number of blocks, And the object cluster and the background cluster according to the color average value of the block.

In the present invention, the automatic object segmentation method using Block Clustering based on GrabCut and the detailed configuration of the automatic object segmentation system 10 will be described in detail.

That is, the object region estimation step S300 includes a cluster labeling processing step S310 for performing labeling processing on the background information, neutral cluster, object cluster, and object region estimation on the image information.

The object region estimation step S300 includes a block clustering process S320 of dividing the image including the object region information into a predetermined number of blocks and estimating the ROI of the object in the block-divided image And the like.

The object interest region estimating unit 30 for estimating the object region of interest (S300) includes a background labeling processing unit 30 for performing labeling processing on the image information, a neutralization cluster, And a block clustering processor 32 for dividing an image including the information of the object region into a predetermined number of blocks and estimating a region of interest of the object in the block-divided image. The cluster labeling processing step S310 is performed by the cluster labeling processing unit 31 and the block clustering processing step S320 and the like are performed by the block clustering processing unit 32. [

The cluster labeling processing step S310 includes a cluster distributed information analysis step S311 for analyzing the distributed information of each cluster for labeling the background of the image, the neutral image, the analyzed background cluster, the neutral cluster, and the object cluster, And an outline information analysis step S312 for analyzing the outline area information of the image.

In the analysis step of the cluster distributed information (S311), a cluster having a large dispersed area is discriminated as a background cluster among the background cluster, a neutral cluster, and an object cluster, and a cluster having a small dispersed area is identified as an object cluster. In addition, the cluster labeling processing unit 31 receives distribution information of each cluster to label the background of the image, the neutral, the analyzed background cluster, the neutral cluster, and the object cluster by the processing of the cluster dispersion information analysis step S311 A cluster distributed information analysis unit 311 for analyzing the cluster distributed information.

In step S311, the distributed information of each cluster may be calculated by Equation (2).

(2)

,

,

와

는 두 축의 평균값).
('area' is the total area of the image (class), x and y are the axes in the class, C is the class,

Wow

Is the mean value of the two axes).

In addition, the edge information analyzer 312 may be provided to analyze the edge region information of the image by processing the edge information analysis step S312. Accordingly, the outline information analysis step S312 may be performed to calculate the outline information of each of the clusters using Equation (4).

(Equation 4)

,

,

(G is each Gaussian area, and τ is the total boundary region).

In addition, a neutral community labeling step (S313) of labeling the neutral community with a background community or an object community according to Equation (5) will be performed. The neutral cluster labeling unit 313 performs labeling processing on the neutral clusters as a background cluster or an object cluster for the performance of the neutral cluster labeling step S313.

(5)

는 중립영역의 분산,

는 외곽영역 안에서 중립영역이 차지하고 있는 면적값,

와

는 각각 객체의 분산과 배경의 분산,

와

는 각각 외곽영역 안에 객체와 배경이 차지하고 있는 면적값).
(foreground is an object cluster, background is a background cluster,

The dispersion of the neutral region,

Is the area value occupied by the neutral region in the outer region,

Wow

And the distribution of the background,

Wow

Is the area value occupied by the object and the background in the outer area).

The block clustering process S320 of dividing the image including the object region information into a predetermined number of blocks and estimating the region of interest of the object in the block-divided image in the object region estimation step S300 (S321) for calculating an area value of a block corresponding to an object cluster in an individual block, a background area value in a block corresponding to a background cluster in an individual block, A block color information calculation step (S322) of calculating a distance between the average color value of each block and the average color value of the background cluster, a block color information calculation step (S322) of calculating a block color area average value, By comparing the object area, the background area, and the color value in the individual block calculated by the calculation step S322, the individual block is classified into the object cluster or the background cluster, It is to perform, including the labeling step (S323) or the like.

The block clustering processing unit 32 for dividing the image including the object region information into a predetermined number of blocks and estimating the ROI of the object in the block-divided image among the object ROI estimation units 30, A block area calculation unit 321 for calculating an object area value in a block corresponding to an object cluster, a background area value in a block corresponding to a background cluster in an individual block, and an area calculation unit 321 for calculating an average color value A block color information calculation unit 322 for calculating a distance between the block color information and an average color value of the background cluster with respect to a color average value for each individual block and a block color information calculation unit 322 for calculating a block color area calculation unit 321 and a block color information calculation unit 322 And an individual block labeling unit 323 for labeling individual blocks as object clusters or background clusters by comparing object areas, background areas, and color values in individual blocks calculated by the individual block It will cost.

The block area calculation step S321 calculates an object area value and a background area value for an individual block according to Equation (6).

(6)

는 개별 블록 중에서 객체의 면적값,

는 개별 블록 중에서 배경의 면적값,

는 객체의 면적값,

는 배경의 면적값,

는 분할된 영상의 개별 블록),
(

Is an area value of an object among individual blocks,

The area value of the background in the individual blocks,

Is the area value of the object,

Is the area value of the background,

Is an individual block of the divided image),

The block color information calculation step S322 calculates an average color value of each of the object cluster and the background cluster with respect to the color average value of the individual blocks using Equations (7) and (8).

(7)

,

,

는 개별 블록의 평균색상과 객체의 평균색상의 거리,

는 개별 블록의 평균색상과 배경의 평균색상의 거리,

는 각 블록의 평균색상,

는 객체의 평균색상,

는 배경의 평균색상).
(

Is the distance between the average color of the individual blocks and the average color of the object,

Is the distance between the average color of the individual blocks and the background color,

Is the average color of each block,

The average color of the object,

The average color of the background).

(8)

,

,

는 normalization 된 개별 블록과 객체 사이의 평균색상거리,

는 normalization 된 개별 블록과 배경 사이의 평균색상거리).
(

Is the average color distance between the normalized individual blocks and objects,

Is the average color distance between the individual blocks normalized and the background).

In the individual block labeling step S323, an individual block is labeled with an object cluster or a background cluster according to Equation (9).

(9)

The object cluster and the background cluster are identified and labeled by performing the block-by-block area calculation step S321, the block color information calculation step S322, and the individual block labeling step S323 of the block clustering processing step S320, The region of interest is estimated. The GrabCut processing unit 40 receiving the estimated image information of the ROI of the object performs the GrabCut processing step S400 to calculate the result image information of the object as shown in the right picture of FIG.

The automatic object segmentation method and the automatic object segmentation system 10 using GrabCut based Block Clustering according to the present invention will now be described with reference to the accompanying drawings.

First, as shown in FIG. 1, an initial segmentation process (object information analysis step) receives image information composed of RGB colors. Then, the K-means algorithm and the GMM algorithm are used to obtain approximate information about the object. Then, in the automatic initial region of Interest Estimation using block clustering process, the cluster region represented by the object and the background is labeled using the analysis information obtained in the previous process, The region of interest is calculated. Finally, in the Object Segmentation process (GrabCut processing step), information of the object region of the image object obtained in the previous process is received and information of the object segmented from the background is obtained by the GrabCut algorithm.

1) Initial Segmentation

First, the process of analyzing the object information (S200, Initial Segmentation) will be described. This object information analysis step is for obtaining approximate information about an object (background) in an input image.

a. K-means clustering

The K-means clustering algorithm (K-means clustering step) is a representative clustering algorithm that labels similar components even if there is no information about the objects. Here, the K-means clustering algorithm is used to estimate the initial parameters (average, covariance, etc.) of the GMM algorithm to be used in the next part.

b. GMM clustering

The GMM (Gaussian Mixture Model) clustering algorithm is a density estimation method that improves the modeling of the distribution density of a given sample data set as a single probability density function. It is a method of modeling. The GMM having the total probability density function M is expressed by Equation (1).

(i = 1, ......, M)는 자료 x에 대해서

(평균벡터) 째 성분 파라미터

(공분산)로 이루어진 확률밀도함수를 의미하며,

는 혼합 가중치(mixture weight)이다.Where x is the D-dimensional (R, G, B) continuous-valued data,

(i = 1, ......, M) is for data x

(Mean vector) th component parameter

(Covariance) of the probability density function,

Is the mixture weight.

In the present invention, the GMM algorithm divides the input image into three clusters represented by "background cluster", "neutral cluster" and "object cluster" using the initial values of GMM obtained by the K-means clustering process, The EM (Expectation - Maximization) algorithm is used. FIG. 2 shows an image divided into three clusters by this GMM processing step.

2) Automatic Initial Region of Interest Estimation using Block clustering

This Automatic Initial Region of Interest Estimation using block clustering is performed by using the result of the previous process (initial segmentation) and setting the region of interest around the image object. The result of this process is used to automatically draw the next area of interest, GrabCut.

a. Labeling class

In this labeling class process (cluster labeling process step), the result of GMM processing that divided the input image into three clusters in the initial segmentation (object information analysis step) is labeled as object cluster, background cluster, neutral cluster will be. To do this, we analyze the outer region information of the image together with the variance information of each cluster.

First, by using the distributed information of classes including each cluster, the background cluster becomes larger than the object cluster by analyzing the distributed information (cluster distributed information analysis step). Next, by performing the check of the outer region information, the cluster portion corresponding to the background cluster is distributed in the outer region, so that the background cluster can be analyzed by analyzing the outer region information (outer information analysis step).

Equation (2) is an equation for obtaining the variance information by analyzing the pixel information of each cluster.

와

는 두 축의 평균값이 된다.
Where 'area' is the total area of the image (class), x and y are the axes in the class, C is the class,

Wow

Is the average of the two axes.

Equation (3) is used to normalize the variance of each cluster obtained from Equation (2).

According to the processing of the variance value, the variance value of the background cluster becomes large whereas the smallest variance value will be discriminated as the object cluster, so that the labeling process is performed as "background cluster" and "object cluster" .

Next, equation (4) is used to determine the outer area information.

Where G is each Gaussian area, and τ is the total boundary region.

Then, the size of the foreground (object group) and the background (background) is compared with each other, and the input pixel is labeled toward the smaller calculation result. This is expressed by Equation (4). Equation (4) is to measure how much each cluster occupies an area in the outer area.

The cluster having the largest variance value is labeled as the "background cluster " and the cluster having the smallest value is labeled as the" object cluster "using the information obtained by the above Equations 2 to 4, The median value is calculated as a "neutral cluster ".

The next step is to label the neutral clusters as object clusters, unless they are background clusters (neutral clustering labeling). Equation (5) calculates the Euclidean distance between the background cluster and the object cluster obtained for the neutral cluster, and labels the cluster closer to the closer one.

는 중립영역의 분산,

는 외곽영역 안에서 중립영역이 차지하고 있는 면적값,

와

는 각각 객체의 분산과 배경의 분산,

와

는 각각 외곽영역 안에 객체와 배경이 차지하고 있는 면적값이다.
Here foreground is the object cluster, background is the background cluster,

The dispersion of the neutral region,

Is the area value occupied by the neutral region in the outer region,

Wow

And the distribution of the background,

Wow

Is the area value occupied by the object and the background in the outer area, respectively.

FIG. 3 is an exemplary view illustrating a result of object region estimation by labeling a background image, an object image, and a neutral image (background or object) by a cluster labeling process.

b. Block clustering

In this block clustering process, region information of an object cluster is processed by dividing and processing an image estimated in the previous process (cluster labeling process step) into a predetermined number of blocks, and then, in processing of the GrabCut, .

First, the input image is divided into a predetermined number of blocks. In the exemplary embodiment of the present invention, the number of blocks may be divided into 10 * 10, i.e., 100 blocks. The number of such blocks may be appropriately determined according to the state or capacity of the processed image, There will be.

Each block divided by a predetermined number is labeled with a background cluster and an object cluster using the image processing result information in the previous process (cluster labeling process step, Labeling Class) to set a rectangle (ROI) around the object.

Next, two feature elements are calculated for each block. First, each cluster calculates the area value occupied by each block. The second compares the color averages in the individual blocks with the color averages of the blocks represented by the object clusters and the color averages of the blocks represented by the background clusters. This process is performed by Equations (6) to (8).

는 개별 블록 중에서 객체의 면적값(each Gaussian),

는 개별 블록 중에서 배경의 면적값(each Gaussian),

는 객체의 면적값,

는 배경의 면적값,

는 분할된 영상의 개별 블록이다.here

(Each Gaussian) of the object in the individual blocks,

Is the area of the background (each Gaussian),

Is the area value of the object,

Is the area value of the background,

Are individual blocks of the segmented image.

That is, Equation (6) calculates how much the object and the background occupy the respective areas in the individual blocks.

는 개별 블록의 평균색상과 객체의 평균색상의 거리,

는 개별 블록의 평균색상과 배경의 평균색상의 거리,

는 각 블록의 평균색상,

는 객체의 평균색상,

는 배경의 평균색상이다.here

Is the distance between the average color of the individual blocks and the average color of the object,

Is the distance between the average color of the individual blocks and the background color,

Is the average color of each block,

The average color of the object,

Is the average color of the background.

We use equation (8) to normalize it.

는 normalization 된 개별 블록과 객체 사이의 평균색상거리,

는 normalization 된 개별 블록과 배경 사이의 평균색상거리가 된다.
here

Is the average color distance between the normalized individual blocks and objects,

Is the average color distance between the normalized individual blocks and the background.

Using Equation (9) which applies the area (Bf, Bg) occupied by the object and the background to the individual blocks and the average color of the object block and the average color of the background block with respect to the average color of the individual blocks, And the corresponding block is labeled with an object cluster and a background cluster.

If the above process is performed, the data of the area of interest (rectangle in the second figure on the right side of FIG. 7) around the final object cluster is calculated, and the blocks of FIG. 4 illustrate the process of the block clustering.

3) Object Segmentation

This Object Segmentation (GrabCut processing step) divides the object using the GrabCut algorithm for the region of interest of the object estimated by the preceding process (object region estimation step). In applying the general GrabCut algorithm applied thereto, it is a powerful algorithm for dividing an object in an image. In the present invention, prior to using the GrabCut algorithm, By applying the automatic object segmentation method according to the present invention, it is possible to automate the process of setting an ROI of an object.

4) Effective experimental example

An example of the image and the process of analyzing by applying the automatic object segmentation method (automatic object segmentation system) using the GrabCut based Block Clustering according to the present invention is as shown in FIG. In addition, images that can be applied to the present invention may use images of various sizes including 250X210 to 450X420.

As an effective example of the present invention, a result obtained by directly drawing a region of interest around an object of an image (original) and a result of setting a region of interest of an object automatically segmented by the present invention are used .

The segmented images obtained by these two methods are compared with the ground truth image and scored using Equation (10). In Equation (10), it can be seen that the score S approaches 1 as the ground truth image and the segmented image become similar.

Table 1 below shows the results of calculation for all images by Equation (10). It is divided into two, average is average and standard deviation is standard deviation. The higher the average, the better the better. The lower the standard deviation, the higher the performance.

method average 표준 deviation Grab Cut (original) 0.94 0.03 proposed method 0.90 0.04

Experimental results show that the performance of the original GrabCut, that is, the performance by the method proposed by the present invention and the conventional implementation in which the region of interest is specified by the user, comes to be almost similar. That is, even if the method according to the present invention is used, it is similar to what the conventional user designates, whereas according to the present invention, it is considered that the efficiency is superior because the automatic processing is performed and processed quickly. Therefore, in the present invention, the entire process is automated and the performance is similar to that of the original GrabCut, so that the performance is actually improved.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. The technical idea of the present invention should not be construed as being limited.

10: automatic object division system 20: object information analysis unit
30: Object ROI estimation unit 31: Cluster labeling processing unit
311: Cluster dispersion information analysis unit 312: Outline information analysis unit
313: Neutral cluster labeling unit 32: Block clustering processing unit
321: block-by-block area calculating unit 322: block color information calculating unit
323: Individual block labeling section
40 GrabCut processing unit

Claims (15)

An object information analysis step of analyzing object information and background information of the image;
An object region estimation step of estimating a region of interest of an object with respect to image information including background and object division information; And
A GrabCut processing step of dividing the object by the GrabCut algorithm using the information of the ROI of the object analyzed by the ROI estimation step;
The method comprising the steps of:
And an object region of interest estimation step of estimating a region of interest of the object with respect to image information including background and object division information.
Wherein the object region of interest estimation comprises:
We analyze clustering information of image and discriminate clusters according to the size of the clustering area,
Wherein an image is divided into a predetermined number of blocks, and an object cluster and a background cluster are determined according to an object area and a background area in an individual block, and a color average value of the block.
3. The method according to claim 1 or 2,
Wherein the object region of interest estimation comprises:
A cluster labeling processing step of performing labeling processing and object area estimation on a background cluster, a neutral cluster, and an object cluster with respect to image information;
The method comprising the steps of:
The method of claim 3,
Wherein the cluster labeling processing step comprises:
A cluster distributed information analysis step for analyzing the distributed information of each cluster to label the background of the image, the neutral, the analyzed background cluster, the neutral cluster, and the object cluster; And
An outer information analysis step of analyzing the outer area information of the image;
The method comprising the steps of:
5. The method of claim 4,
In the cluster dispersion information analysis step, a cluster having a large dispersion area among the background clusters, a neutral cluster, and an object cluster is determined as a background cluster, and a cluster having a small dispersion area is determined as an object cluster.
5. The method of claim 4,
The step of analyzing the cluster distributed information may include: calculating dispersion information of each cluster by Equation (2); And
(2)

,
('area' is the total area of the image (class), x and y are the axes in the class, C is the class,

Wow

Is the mean value of the two axes),
The perimeter information analysis step S312 may include calculating perimeter information of each of the clusters using Equation (4);
(4)

,
(G is each Gaussian area, τ is a total boundary region)
The method comprising the steps of:
The method of claim 3,
A neutral community labeling step of labeling the neutral community with a background community or an object community according to Equation (5);
(5)

(foreground is an object cluster, background is a background cluster,

The dispersion of the neutral region,

Is the area value occupied by the neutral region in the outer region,

Wow

And the distribution of the background,

Wow

Is the area value occupied by the object and the background in the outer area, respectively)
The method comprising the steps of:
3. The method according to claim 1 or 2,
Wherein the object region of interest estimation comprises:
A block clustering processing step of dividing an image including the information of the object area into a predetermined number of blocks and estimating a ROI of the object in the divided images;
The method comprising the steps of:
The method of claim 8,
Wherein the block clustering processing step comprises:
An object area value in a block corresponding to an object cluster in an individual block, a block area calculation step for calculating a background area value in a block corresponding to the background cluster in the individual block;
A block color information calculating step of calculating a distance between the average color value of each block and the average color value of the object cluster, and a distance between the average color value of each block and the average color value of the background cluster; And
An individual block labeling step of labeling individual blocks into object clusters or background clusters by comparing object areas, background areas, and color values in the individual blocks calculated by the block area calculation step and the block color information calculation step;
The method comprising the steps of:
10. The method of claim 9,
The block area calculation step may calculate an object area value and a background area value for each block according to Equation (6)
The block color information calculation step calculates an average color value of each of the object cluster and the background cluster with respect to the color average value of the individual blocks according to Equations (7) and (8)
Wherein the individual block labeling step labels individual blocks as object clusters or background clusters according to Equation (9).

(6)

,
(

Is an area value of an object among individual blocks,

The area value of the background in the individual blocks,

Is the area value of the object,

Is the area value of the background,

Is an individual block of the divided image),
(7)

,
(

Is the distance between the average color of the individual blocks and the average color of the object,

Is the distance between the average color of the individual blocks and the background color,

Is the average color of each block,

The average color of the object,

The average color of the background),
(8)

,
(

Is the average color distance between the normalized individual blocks and objects,

Is the average color distance between the individualized blocks normalized and the background),
(9)

.
3. The method according to claim 1 or 2,
The object information analysis step includes:
A K-means clustering step for performing a clustering algorithm on an image targeted by the K-means algorithm; And
A GMM processing step of dividing an image into a "background", a "neutral", and an "object" as the target image by performing the GMM algorithm according to the mixture weight;
The method comprising the steps of:
An object information analyzer for analyzing object information and background information of an image;
An object region of interest estimating an area of interest of an object with respect to image information including background and object division information; And
A GrabCut processing unit for dividing the object by the GrabCut algorithm using information of the ROI of the object analyzed by the ROI estimating unit;
Lt; / RTI >
The object-of-
We analyze clustering information of image and discriminate clusters according to the size of the clustering area,
Wherein the image is divided into a predetermined number of blocks and the object cluster and the background cluster are discriminated according to the object area and the background area in the individual block and the color average value of the block.
13. The method of claim 12,
Wherein the object ROI estimator comprises:
A community labeling processor for performing a labeling process and an area estimation of an object on a background cluster, a neutral cluster, and an object cluster with respect to image information; And
A block clustering processing unit for dividing an image including information of an object area into a predetermined number of blocks and estimating a ROI of the object in the divided images;
And an automatic object segmentation system.
14. The method of claim 13,
Wherein the cluster labeling processing unit comprises:
A cluster distributed information analysis unit for analyzing the distributed information of each cluster to label the background of the image, the neutral, the analyzed background cluster, the neutral cluster, and the object cluster;
An outer information analyzing unit for analyzing the outer area information of the image; And
A neutral community labeling unit for labeling the neutral community with a background community or an object community;
And an automatic object segmentation system.
14. The method of claim 13,
Wherein the block clustering processor comprises:
A block area calculation unit for calculating an object area value in the block corresponding to the object cluster in the individual block and a background area value in the block corresponding to the background cluster in the individual block;
A block color information calculation unit for calculating a distance between a color average value of each block and an average color value of the object cluster and a color value average of each individual block and an average color value of the background group; And
An individual block labeling unit for labeling individual blocks as object clusters or background clusters by comparing object areas, background areas, and color values in individual blocks calculated by the block area calculation unit and the block color information calculation unit;
And an automatic object segmentation system.

Images