KR101934358B1 - Method and Apparatus for Image Segmentation Based on Multiple Random Walk - Google Patents

Method and Apparatus for Image Segmentation Based on Multiple Random Walk Download PDF

Info

Publication number
KR101934358B1
KR101934358B1 KR1020150162823A KR20150162823A KR101934358B1 KR 101934358 B1 KR101934358 B1 KR 101934358B1 KR 1020150162823 A KR1020150162823 A KR 1020150162823A KR 20150162823 A KR20150162823 A KR 20150162823A KR 101934358 B1 KR101934358 B1 KR 101934358B1
Authority
KR
South Korea
Prior art keywords
random
image
random walk
node
graph
Prior art date
Application number
KR1020150162823A
Other languages
Korean (ko)
Other versions
KR20170058779A (en
Inventor
박병관
김창수
장원동
이세호
김영배
Original Assignee
에스케이 텔레콤주식회사
고려대학교 산학협력단
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 에스케이 텔레콤주식회사, 고려대학교 산학협력단 filed Critical 에스케이 텔레콤주식회사
Priority to KR1020150162823A priority Critical patent/KR101934358B1/en
Publication of KR20170058779A publication Critical patent/KR20170058779A/en
Application granted granted Critical
Publication of KR101934358B1 publication Critical patent/KR101934358B1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/12—Edge-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10024—Color image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10028—Range image; Depth image; 3D point clouds

Abstract

A random walk-based image segmentation method and apparatus based on a multiple random walk is disclosed.
The result of the random walker distribution is obtained by dividing the image including the color (RGB) information and the depth information into superpixels and applying a multiple random walk based on the superpixel-based graph And more particularly, to a method and apparatus for multi-random walk based image segmentation.

Figure R1020150162823

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to an image segmentation method and apparatus for multi-

The present embodiment relates to a method and an apparatus for dividing an image based on a multiple random walk.

The contents described in this section merely provide background information on the present embodiment and do not constitute the prior art.

In general, techniques for image segmentation include a graph-based image segmentation technique using color information (RGB) (hereinafter referred to as 'first prior art') or a graph-based image segmentation technique using color information and depth information (Hereinafter referred to as " second prior art ").

The first conventional technique performs graph-based image segmentation using color information. The first conventional technique connects an edge between neighboring pixels using color information, and detects a minimum weight cut that minimizes a weight sum of edges in an image in an image, thereby performing image segmentation. The first conventional technique can divide the image into two regions such that the color difference between the regions is maximized, and the user can manually divide the region of the two regions to set an initial condition, .

The second prior art is a graph-based image segmentation technique using color information and depth information. 2, the depth histogram is analyzed to find a boundary line 110, then an offset is applied to calculate two threshold values, and a threshold value for each of the foreground region and the background region is used Foreground region, background region and remaining unknown region. The second conventional technique performs final image segmentation by connecting pixels of an unknown region with edges and performing graph-cutting using a minimum weight to classify the unknown region into a foreground region and a background region. The second conventional technique effectively classifies the unknown area into the foreground area and the background area by minimizing the weight. However, when the threshold value is determined incorrectly in the initial process, the foreground and background areas are wrong And the final image segmentation fails.

In this embodiment, an image including color (RGB) information and depth information is divided into superpixels, and a multiple random walk is applied based on a superpixel-based graph. The present invention provides a method and an apparatus for dividing an image according to a result of a probability.

According to an aspect of the present invention, there is provided a super pixel division method for dividing an image including color (RGB) information and depth information into a super pixel; Setting a superpixel as a node, and connecting each node with an edge based on color information and depth information of a superpixel set in each node to construct a graph; A random walk process for calculating a distribution probability of at least one random walker among the nodes by applying a multiple random walk based on the graph and calculating a final normal distribution for the random walker ; And assigning a label to the random walker to the node based on the final normal distribution and dividing the image based on the label. Provides a partitioning method.

According to another aspect of the present invention, there is provided a super pixel division unit for dividing an image including color information and depth information into super pixels; Setting a super pixel as a node, connecting each node to an edge based on color information and depth information of a super pixel set at each node, and constructing a graph; A random walk processing unit for calculating a distribution probability of at least one random walker among the nodes by applying a multiple random walk based on the graph and calculating a final normal distribution for the random walker; And a video dividing unit for assigning a label for the random walker to the node based on the final normal distribution and dividing the image based on the label. to provide.

As described above, according to the present embodiment, the multi-random walk-based image segmentation device can divide an image into units having similar characteristics, and can improve the segmentation performance compared to the case of using only images containing only color information There is an effect.

In addition, the multi-random work-based image segmentation device can be utilized for accurate image segmentation in future object recognition, and can be divided into a portable terminal, a robot, a security camera or a general camera by combining with a 3D camera .

FIGS. 1A and 1B are diagrams for explaining general graph-based image segmentation.
FIG. 2 is a diagram schematically illustrating the operation of a random walk-based image segmentation apparatus according to the present embodiment.
3 is a block diagram schematically showing the detailed structure of the random walk-based image segmentation device according to the present embodiment.
4 is a flowchart illustrating a random walk-based image segmentation method according to the present embodiment.
5 is an exemplary diagram for explaining an operation of dividing an image into super pixels in the image dividing device according to the present embodiment.
6 is an exemplary diagram illustrating a structure of a graph having a multi-layer structure according to the present embodiment.
7 is a diagram illustrating an input image and a random walk based image segmentation result image according to the present embodiment.
8 is a diagram illustrating an example of a random walk based image segmentation result image according to the present embodiment.

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

Recently, intelligent systems have been actively researched, for example, intelligent automobiles or humanoid robots are representative intelligent systems. These intelligent systems recognize the surrounding space and identify the person or detect the object.

In order to perform spatial recognition, accurate analysis of 3D spatial information must be preceded. Image segmentation is a typical image analysis technique that divides an image into objects using color information and depth information of the input image, and can be a base technology for various applications.

Low-level processing, which increases the processing efficiency of algorithms linked to existing image segmentation by reducing the number of pixels, is focused on. In recent years, however, information about objects existing in the image, including object detection and recognition, Which is an object-based image segmentation. Accordingly, the embodiment of the present invention proposes a technique of segmenting an RGB-D image on an object basis.

The image segmentation apparatus according to the present embodiment probabilistically performs image segmentation using a multiple random walk instead of the graph cut used in the first prior art and the second prior art.

Since the image dividing apparatus according to the present embodiment uses a multi-random walker, it is possible to divide a desired number of regions, and there is no need to set initial threshold values or initial conditions for image segmentation, There is a difference from the first prior art and the second prior art in that it increases the cost.

FIG. 2 is a diagram schematically illustrating the operation of a random walk-based image segmentation apparatus according to the present embodiment.

The image segmenting apparatus 200 according to the present embodiment performs an operation 210 of acquiring an image including color (RGB) information and depth information from a camera.

The image dividing device 200 divides an image obtained by a super pixel unit by applying a super pixel technique and then sets an super pixel as a node and connects an edge to an edge to construct a graph 220 .

(Multiple Random Walk) based on the graph, and performs an operation 230 of dividing the image according to a result of the probability that a random walker is distributed in each node.

3 is a block diagram schematically showing the detailed structure of the random walk-based image segmentation device according to the present embodiment.

The image dividing apparatus 200 according to the present embodiment includes an image obtaining unit 310, a super pixel dividing unit 320, a graph forming unit 330, a random walk processing unit 340 and an image dividing unit 380 . 3 is one embodiment of the present invention. All the blocks shown in FIG. 3 are not essential components. In another embodiment, some of the blocks included in the image segmentation apparatus 200 are added , Changed or deleted.

The image acquiring unit 310 acquires an image including color (RGB) information and depth information from a camera (not shown). The color information included in the acquired image means information including three color-based color values of Red (Red), Green (Green), and Blue (Blue). The depth information is information including a depth value indicating how far the pixels of the image are relative to the camera.

The image acquired by the image acquiring unit 310 may be an RGB-D image, but it may be another type of image if it includes information related to color and depth.

The super pixel division unit 320 divides the acquired image into super pixels by applying a predetermined technique to generate a super pixel divided image. The super pixel division unit 320 divides the pixels included in the image into a small uniform region having similar characteristics, that is, a super pixel which is a basic unit for image processing that does not include an edge.

The super-pixel division unit 320 may divide an image into super-pixels by applying a graph-based technique or a gradient-based technique. The graph-based technique constructs a graph by setting each pixel of the image as a node of the graph and assigning the feature between the pixel and the pixel as the edge value of the graph. It means a method of obtaining the super-pixel by dividing the graph into two sub-graphs repeatedly by obtaining the eigenvector and the eigenvalue from the N × N weighting matrix for all the nodes of the graph. The slope-based technique calculates the slope value of an image, sets an initial pixel (Seed) based on the slope value, calculates a Euclidean distance between the pixel and the initial pixel, divides the pixel into a small area having similar properties, . Here, the gradient-based technique may be a mean shift (MS), a turbo pixel (Turbopixel), a quick shift, a simple linear Iterative Clustering (SLIC), or the like.

The super-pixel division unit 320 can divide an image into super-pixels using one technique, but can divide an image into super-pixels using a plurality of techniques. The super-pixel division unit 320 generates a super-pixel-divided image for each technique when dividing the super-pixel by applying two or more techniques.

The graph configuring unit 330 configures a superpixel as a node and connects each node to an edge based on color information and depth information of the superpixel set at each node to construct a graph.

The graph configuring unit 330 may configure each superpixel as a node and connect edges to form a bisection graph. E (Edge) is an edge (edge) that connects different nodes. The edge (edge) is the edge that connects the different nodes. ). In other words, this graph divides the vertex set V into two sets V 1 and V 2 , and all sides (e) included in the set E can be simultaneously tangent to the vertex of V 1 and the vertex of V 2 Means a graph.

In the case of constructing a graph using a plurality of super-pixel divided images by using a plurality of techniques, the graph configuring unit 330 configures a graph having a multi-layer structure by setting each super-pixel divided image as a different layer . For example, in the case of two super-pixel divided images, the graph configuring unit 330 sets one super-pixel divided image as an upper layer and sets the remaining super-pixel divided images as a lower layer, You can construct a graph.

The graph composing unit 330 connects edges between nodes in a graph having a multi-layer structure, and connects nodes at the same position to the edges when connecting nodes included in different layers. On the other hand, when connecting the nodes included in the same layer, the graph configuring unit 330 connects adjacent nodes with edges, but connects neighboring nodes only in the upper layer and does not connect the nodes in the lower layer.

The graph constructing unit 330 assigns weights of edges based on difference values between color information and depth information between different nodes.

Hereinafter, in order to explain the operation for determining the weights to be given to the edges in the graph constructing unit 330, weights w (i, j) of edges e ij connected between arbitrary first node i and second node j ij ), as shown in FIG.

The graph constructing unit 330 determines the weight w ij of the edge e ij connected between the first node i and the second node j using Equation 1.

Figure 112015113165412-pat00001

In Equation (1), w ij denotes the weight of the edge e ij , c i and d i mean the average Lab color value and depth value of the first node i , respectively, and c j and d j Mean the average Lab color value and depth value of the second node j, respectively. Further, in Equation (1)

Figure 112015113165412-pat00002
And
Figure 112015113165412-pat00003
Means a parameter indicating a predetermined weight for color information and depth information, respectively.

The weight w ij of the edge e ij in the graph composing unit 330 is determined by the first node i and the second node j based on the color information and the depth information of the first node i and the second node j, (j), the smaller the difference in color and depth, the higher the weight.

The random walk processing unit 340 performs a plurality of random walker based image segmentation by applying a multiple random walk. In other words, the random walk processing unit 340 performs image segmentation by causing a plurality of random walkers to move on one graph and have a feature of pushing each other at a predetermined node. Here, the characteristic of pushing each other means that, when the distribution probability of a predetermined random walker at a predetermined node is high, the remaining random walkers except a predetermined random walker have a low distribution probability at a predetermined node.

The random walk processing unit 340 calculates a distribution probability of at least one random walker among the nodes by applying a multiple random walk based on the graph and calculates a final normal distribution for the random walker . Here, the random walk refers to a random walk model that utilizes the probability of moving to a different node in an arbitrary direction in a predetermined node, and the random walker refers to a random walk The probability distributions on the graphs for the graphs.

The random walk processing unit 340 according to the present embodiment includes a transition matrix generation unit 350, a single random walk processing unit 360, and a multiple random walk processing unit 370.

The transition matrix generator 350 calculates a transition probability that the random walker will transition from the second node j to the first node i, and generates a transition matrix using the transition probability.

The transition matrix generator 350 calculates the transition probability a ij to be transferred from the second node j to the first node i at a weight w ij of the edge e ij , j) is calculated by dividing the degree. Here, the order of the second node j means the number of edges including the second node j.

The transition matrix generator 350 calculates the transition probability that the random walker will transition from the second node j to the first node i using Equation (2).

Figure 112015113165412-pat00004

In Equation (2), a ij denotes a probability that the random walker will transition from the second node (j) to the first node (i), and w ij denotes a weight of the edge (e ij ).

The transition matrix generator 350 generates the transition matrix A using the transition probability a ij . Here, the transition matrix A is composed of A = [a ij ].

The single random walk processing unit 360 performs an operation of determining one or more random walkers before performing the multiple random walk in the multiple random walk processing unit 370. [

The single random walk processing unit 360 performs a single random walk using the transition matrix A to calculate a stationary distribution. Here, the single random walk processor 360 performs a single random walk on the assumption that one random walker exists, and calculates a normal probability distribution using the random walker's distribution probability for each of the plurality of nodes. Here, the normal probability distribution means a distribution of the probability that a random walker moves to a predetermined node.

The single random walk processing unit 360 determines an initial distribution of at least one random walker for applying the multiple random walk based on the size of the normal probability distribution. Here, it is preferable that the number of the random walkers is previously set by the setting of the user. For example, the single random walk processing unit 360 sets a node having the highest value in the normal probability distribution as an initial node (Seed) for applying the multi-random work, that is, as a random walker, do.

If the distribution of the newly determined random walkers is lower than a preset threshold value, the single random walk processing unit 360 repeats the operation of calculating the normal probability distribution again and setting the random walker.

The multiple random walk processing unit 370 calculates a distribution probability of at least one random walker among the nodes included in the graph by applying a Multiple Random Walk and calculates a final normal distribution for a plurality of random walkers . In other words, the multi-random walk processor 370 calculates the distribution probability of each of the multiple random walkers at a predetermined time, and calculates the final normal distribution for each multi-random walker based on the distribution probability.

The multi-random walk processing unit 370 calculates the distribution probability of the random walker (k) at time t

Figure 112015113165412-pat00005
) Is determined using Equation (3).

Figure 112015113165412-pat00006

In Equation (3)

Figure 112015113165412-pat00007
(K) denotes the distribution probability of the random walker (k) at time t,
Figure 112015113165412-pat00008
Denotes a predetermined restart probability, and K denotes the total number of random walkers. Also,
Figure 112015113165412-pat00009
(Restart vector) in the random walker k.

The multi-random walk processing unit 370 calculates a random walker (k) distribution probability

Figure 112015113165412-pat00010
), The restart value (
Figure 112015113165412-pat00011
) To implement mutual pushing between random walkers at a predetermined node. Here, the characteristic of pushing each other means that, when the distribution probability of a predetermined random walker at a predetermined node is high, the remaining random walkers except a predetermined random walker have a low distribution probability at a predetermined node.

The multi-random walk processing unit 370 uses the equation (4) to calculate the restart value (

Figure 112015113165412-pat00012
).

Figure 112015113165412-pat00013

In Equation (4)

Figure 112015113165412-pat00014
(K) < / RTI > at the first node < RTI ID = 0.0 &
Figure 112015113165412-pat00015
(K) at the first node (i) and the distribution probability of the random walk
Figure 112015113165412-pat00016
Denotes a probability of transition from the second node j to the first node i.

That is, the multi-random walk processing unit 370 calculates the distribution probability of the random walker (k) using [Expression 3] and [Expression 4] , The remaining random walkers except for the random walker (k) have a low restart value at the first node (i). The multi-random walk processing unit 370 has the feature of pushing each random walker by allocating the restart value using [Equation 3] and [Equation 4].

The multiple random walk processing unit 370 calculates the final normal distribution of the distribution probability for each of the plurality of random walkers. The multi-random walk processing unit 370 calculates the final normal distribution using [Equation (5)].

Figure 112015113165412-pat00017

In Equation (5)

Figure 112015113165412-pat00018
(K) denotes the distribution probability of the random walker (k).

The multi-random walk processing unit 370 calculates a convergence value of the final normal distribution for each random walker using Equation (5), and determines a random walker having the highest distribution probability at a predetermined node. For example, when the first to third random walkers are present, the multi-random walk processor 370 sets the first random walker having the highest distribution probability in the nodes included in the predetermined area, The second random walker having the highest distribution probability can be set to the nodes included in the predetermined region and the third random walker having the highest distribution probability can be set to the nodes included in another predetermined region.

The image dividing unit 380 assigns a label corresponding to the random walker determined by the multiple random walk processing unit 370 to each of the plurality of nodes, and divides the image based on the label. For example, when the first to third random walkers are present, the image dividing unit 380 assigns a red label to the first random walker, a yellow label to the second random walker, For walkers, a black label can be assigned. The image segmentation unit 380 may divide a plurality of nodes having the same label among a red label, a yellow label, and a black label into one area. The image divider 380 may display a boundary image between the divided regions of each label to output a result image in which the images are divided.

4 is a flowchart illustrating a random walk-based image segmentation method according to the present embodiment.

The image dividing device 200 acquires an image including color (RGB) information and depth information from a camera (S410), and divides the acquired image into super pixel units (S420). The image segmenting device 200 may divide an image into super pixels by applying a graph-based technique or a gradient-based technique. For example, the image segmenting device 200 may include a mean shift (MS), a turbo pixel (Turbopixel), a quick shift ), SLIC (Simple Linear Iterative Clustering), or the like, each of which can be divided into super-pixel units.

The image dividing device 200 generates a graph based on the color information and the depth information of the super pixel (S430). More specifically, the image dividing device 200 sets a super pixel as a node, and constructs a graph by connecting each node with an edge based on color information and depth information of a super pixel set as a node. Here, when the super-pixel is divided by using a plurality of techniques, the image dividing device 200 can configure the super-pixel divided image according to each technique as a different layer to construct a graph having a multi-layer structure have.

The image dividing device 200 assigns weights based on difference values between color information and depth information between nodes in an edge connecting nodes.

The image dividing device 200 generates a transition matrix for performing the multiple random walk (S440). The image dividing device 200 calculates a transition probability that the random walker will transition from the second node j to the first node i and generates a transition matrix using the calculated transition probability.

The image dividing device 200 performs a single random walk to determine an initial distribution of a random walker for performing the multiple random walk (S450). The image dividing device 200 calculates a normal probability distribution by performing a single random walk using the transition matrix A and calculates a seed value of a seed having a highest value in a normal probability distribution as an initial node Seed ) That is, it is set as a random walker to determine the initial distribution of the multiple random walk. Here, the step S450 of setting the random walker to the initial distribution of the multiple random walk may be omitted when the user sets the initial distribution of the multiple random walk.

The image dividing device 200 calculates the final normal distribution of the random walkers among the nodes included in the graph by applying the multiple random walk (S460). In other words, the image dividing device 200 calculates the distribution probability of each of the multiple random walkers at a predetermined time, and calculates the final normal distribution for each of the multiple random walkers based on the distribution probability. The image dividing device 200 determines a random walker having the highest distribution probability at a predetermined node based on the convergence value of the final normal distribution for each random walker.

The image dividing device 200 assigns a label corresponding to the determined random walker to each of the plurality of nodes, and divides the image based on the label (S470).

5 is an exemplary diagram for explaining an operation of dividing an image into super pixels in the image dividing device according to the present embodiment.

The image dividing device 200 divides the image 510 obtained in order to perform image division based on the multiple random walk into super pixels. The image segmentation apparatus 200 may divide the image 510 by applying a plurality of super pixel techniques.

FIG. 5A shows a first super-pixel-divided image 520 obtained by dividing an image 510 obtained by applying a SLIC (Simple Linear Iterative Clustering) method into super-pixel units.

5B shows a second superpixel segmented image 530, a third superpixel segmented image 532, and a fourth superpixel segmented image 534, which are divided into superpixel units by applying an averaging technique . Here, in applying the average moving method, the second super-pixel divided image 530, the third super-pixel divided image 532, and the fourth super-pixel divided image 534 change the set parameters, .

6 is an exemplary diagram illustrating a structure of a graph having a multi-layer structure according to the present embodiment.

The image dividing device 200 forms a graph by setting each super pixel as a node. 6, the image dividing apparatus 200 includes a first super-pixel division image 520, a second super-pixel division image 530, a third super-pixel division image 532, and a fourth super- The image 534 is used to construct a multi-layered graph.

For example, the image segmentation apparatus 200 sets the nodes included in the first super-pixel segmented image 520 divided into super pixels by using the SLIC technique as an upper layer, applies an average shift technique, The nodes included in the divided second super pixel divided image 530, the third super pixel divided image 532, and the fourth super pixel divided image 534 can be set as a lower layer.

The image dividing device 200 connects edges between nodes in a graph having a multi-layer structure, and includes nodes included in the first super-pixel divided image 520 and second through fourth super-pixel divided images 530, 532, 534, when connecting nodes included in different layers, nodes at the same position are connected to the edge.

Meanwhile, the image dividing device 200 connects the adjacent nodes included in the upper layer, such as the nodes included in the first super-pixel divided image 520, with edges, and the second through fourth super-pixel divided images 530, 532, and 534, nodes included in each of the lower layers are not connected to each other.

FIG. 7 is a diagram illustrating an input image and a random walk based image segmentation result image according to the present embodiment, and FIG. 8 illustrates an example of a random walk based image segmentation result image according to the present embodiment.

7, the image segmenting apparatus 200 includes a first segmentation result image 730 segmented by applying a multi-random walk based on an image including color information 710 and depth information 720, Lt; / RTI > 8 shows a plurality of images 810 to 860 obtained by dividing an image by applying a multiple random walk to various images.

The foregoing description is merely illustrative of the technical idea of the present embodiment, and various modifications and changes may be made to those skilled in the art without departing from the essential characteristics of the embodiments. Therefore, the present embodiments are to be construed as illustrative rather than restrictive, and the scope of the technical idea of the present embodiment is not limited by these embodiments. The scope of protection of the present embodiment should be construed according to the following claims, and all technical ideas within the scope of equivalents thereof should be construed as being included in the scope of the present invention.

As described above, the present embodiment is applied to an image segmentation related field, and the image can be divided into units having similar characteristics, and the segmentation performance can be improved as compared with the case where only the image including only color information is used. It is a useful invention that generates effects that can be utilized for accurate segmentation of images.

200: image dividing device
310: image acquisition unit 320: super pixel division unit
330: Graph forming unit 340: Random walk processing unit
350: transition matrix generation unit 360: single random walk processing unit
370: a multi-random walk processing unit 380:

Claims (8)

A super pixel division step of generating two or more super pixel divided images obtained by dividing the image including color (RGB) information and depth information into super pixels by applying at least two or more techniques to each image;
The super pixels are set as nodes, the nodes are connected to each other based on color information and depth information of the super pixels set to each node, and the two or more super pixel divided images are set as different layers A graph construction process for constructing a graph having a multi-layer structure;
A random walk process for calculating a distribution probability of at least one random walker among the nodes by applying a multiple random walk based on the graph and calculating a final normal distribution for the random walker ; And
Assigning a label for the random walker to the node based on the final normal distribution, and dividing the image based on the label,
Wherein the method comprises the steps of: delete The method according to claim 1,
In the graph construction process,
Connecting the nodes corresponding to each of the superpixels with the edge, connecting nodes at the same position in the different layers with the edge, and connecting the adjacent nodes in the same upper layer with the edge, Multi - random walk based image segmentation method. The method according to claim 1,
In the graph construction process,
Wherein the graph is configured by weighting the edge based on difference values between the color information and the depth information between the different nodes. The method according to claim 1,
The random walk process includes:
A transition matrix generation step of calculating a transition probability that a random walk based node will transition from the second node j to the first node i and generating a transition matrix using the transition probability; And
Calculating a distribution probability of each of the random walkers applying the multi-random walk at a predetermined time using the transition matrix, and calculating a final normal distribution for each of the random walkers based on the distribution probability; process
Wherein the method comprises the steps of: 6. The method of claim 5,
The random walk process includes:
A single random walk process for performing a single random walk using the transition matrix and determining an initial distribution of the random walker for applying the multiple random walk based on the size of the normal distribution calculated by the single random walk The method further comprising the steps of: A super pixel division unit for generating two or more super pixel divided images obtained by dividing the image including the color information and the depth information into super pixels by respectively applying at least two techniques to the images;
The super pixel is set as a node, each node is connected to an edge based on color information and depth information of a super pixel set in each node, and the two or more super pixel divided images are set as different layers, A graph constructing unit for constructing a graph having
A random walk processing unit for calculating a distribution probability of at least one random walker among the nodes by applying a multiple random walk based on the graph and calculating a final normal distribution for the random walker; And
Assigning a label for the random walker to the node based on the final normal distribution, and dividing the image based on the label,
And a second random-walk-based image segmentation unit. 8. The method of claim 7,
The random walk processing unit,
A transition matrix generator for calculating a transition probability that a random walk based node will transition from the second node j to the first node i and generating a transition matrix using the transition probability;
A single random walk processor for performing a single random walk using the transition matrix and determining an initial distribution of the random walker for applying the multiple random walk based on the size of the normal distribution calculated by the single random walk; And
A random random walker for calculating a distribution probability of each of the random walkers applying the multiplex random walk at a predetermined time using the transition matrix and calculating a final normal distribution for each random walker based on the distribution probability,
And a second random-walk-based image segmentation unit.
KR1020150162823A 2015-11-19 2015-11-19 Method and Apparatus for Image Segmentation Based on Multiple Random Walk KR101934358B1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
KR1020150162823A KR101934358B1 (en) 2015-11-19 2015-11-19 Method and Apparatus for Image Segmentation Based on Multiple Random Walk

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
KR1020150162823A KR101934358B1 (en) 2015-11-19 2015-11-19 Method and Apparatus for Image Segmentation Based on Multiple Random Walk

Publications (2)

Publication Number Publication Date
KR20170058779A KR20170058779A (en) 2017-05-29
KR101934358B1 true KR101934358B1 (en) 2019-01-02

Family

ID=59053625

Family Applications (1)

Application Number Title Priority Date Filing Date
KR1020150162823A KR101934358B1 (en) 2015-11-19 2015-11-19 Method and Apparatus for Image Segmentation Based on Multiple Random Walk

Country Status (1)

Country Link
KR (1) KR101934358B1 (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108839032A (en) * 2018-06-27 2018-11-20 深圳大图科创技术开发有限公司 A kind of intelligent robot
CN112330697B (en) * 2020-12-04 2024-04-30 中信银行股份有限公司 Image segmentation method and device, electronic equipment and readable storage medium

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2012234528A (en) * 2011-04-29 2012-11-29 Mitsubishi Electric Corp Method for segmenting image into super-pixels

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2012234528A (en) * 2011-04-29 2012-11-29 Mitsubishi Electric Corp Method for segmenting image into super-pixels

Also Published As

Publication number Publication date
KR20170058779A (en) 2017-05-29

Similar Documents

Publication Publication Date Title
CN107274419B (en) Deep learning significance detection method based on global prior and local context
US11120556B2 (en) Iterative method for salient foreground detection and multi-object segmentation
Lee et al. Multiple random walkers and their application to image cosegmentation
US8248410B2 (en) Synthesizing detailed depth maps from images
CN108537239B (en) Method for detecting image saliency target
US20160133027A1 (en) Method and apparatus for separating foreground image, and non-transitory computer-readable recording medium
US9501837B2 (en) Method and system for unsupervised image segmentation using a trained quality metric
EP3073443B1 (en) 3d saliency map
CN104966290B (en) A kind of adaptive weighting solid matching method based on SIFT description
Boiangiu et al. Voting-based image segmentation
JP6389742B2 (en) Image segmentation method, apparatus, and program
KR101934358B1 (en) Method and Apparatus for Image Segmentation Based on Multiple Random Walk
JP6787844B2 (en) Object extractor and its superpixel labeling method
Ückermann et al. Realtime 3D segmentation for human-robot interaction
JP2014010717A (en) Area division device
Jia et al. Dense interpolation of 3d points based on surface and color
Sima et al. An extension of the Felzenszwalb-Huttenlocher segmentation to 3D point clouds
JP6448109B2 (en) Image area dividing apparatus and image area dividing program
Huart et al. Similarity-based and perception-based image segmentation
KR101611466B1 (en) Subsampling-based Clustering Method for fast superpixel segmentation
JP2013061876A (en) Information processing apparatus, control method for information processing apparatus, and program
US10229337B2 (en) Scale-adaptive media item segmentation technique
Wibisono et al. Fusion of color and depth information for image segmentation
CN108109150B (en) Image segmentation method and terminal
Cambra et al. Improving Depth Estimation Using Superpixels.

Legal Events

Date Code Title Description
A201 Request for examination
E902 Notification of reason for refusal
E701 Decision to grant or registration of patent right
GRNT Written decision to grant