KR20080002409A - Device and method for transforming 2-d image into 3-d image - Google Patents

Abstract

An apparatus and a method for converting a two dimensional image into a three dimensional image are provided to convert the two dimensional image into the three dimensional image by considering presence of a motion of a camera, so that adaptive three dimensional image conversion can be possible. A feature point extracting unit(100) which includes a color dividing unit, a labling unit, a contour extracting unit and a feature point gaining unit divides an area of a two dimensional image and extracts a feature point of the divided areas. A camera motion recognizer(102) recognizes a camera motion in the two dimensional image. A motion parallax converter(104) generates a variance image for the two dimensional image through a motion parallax if the camera motion recognizer recognizes no camera motion. A scene structure estimation converter(106) generates a variation image for the two dimensional image through screen structure estimation if the camera motion recognizer recognizes the camera motion. By using the variation images generated in the motion parallax converter or the scene structure estimation converter, an image synthesizer(110) synthesizes a three dimensional image.

Description

Device and method for transforming two-dimensional images to three-dimensional image {Device and Method for transforming 2-D Image into 3-D Image}

1 is a block diagram showing the configuration of a three-dimensional conversion apparatus of a two-dimensional image according to an embodiment of the present invention.

Figure 2 is a block diagram showing the configuration of a feature point extraction unit according to an embodiment of the present invention.

3 is a diagram illustrating an image in which color division is completed according to an embodiment of the present invention.

4 is a diagram illustrating an example of performing labeling on a target image to be converted according to an exemplary embodiment of the present invention.

5 is a diagram illustrating an image after labeling the image of FIG. 3 according to an embodiment of the present invention; FIG.

6 illustrates an example of finding a boundary for contour extraction according to an embodiment of the present invention.

FIG. 7 illustrates an image of completing contour extraction on the labeled image of FIG. 5; FIG.

 FIG. 8 is a diagram illustrating an image from which feature points are extracted from the image of FIG. 7. FIG.

9 is a flowchart illustrating a process of recognizing whether there is a camera panning in a camera motion recognition unit according to an exemplary embodiment of the present invention.

10 is a flowchart illustrating a process of determining a camera zoom in and a camera zoom out according to an exemplary embodiment of the present invention.

11 is a flowchart illustrating a process of generating a disparity image using motion parallax according to an exemplary embodiment of the present invention.

12 is a flowchart illustrating a disparity image generation process using scene structure estimation according to an exemplary embodiment of the present invention.

FIG. 13 illustrates a camera model for scene structure estimation according to an embodiment of the present invention. FIG.

The present invention relates to an apparatus and method for converting a two-dimensional image to a three-dimensional image, and more particularly, to an apparatus and method for converting a two-dimensional image into a three-dimensional image using a disparity image.

In the real world, the eye is automatically focused on what the human is seeing, but since all images displayed on the computer monitor do not have depth cues, the focus is always constant. Therefore, the 3D display artificially provides different images to the left and right, but since the distance from the eye to the object is always constant, a problem arises in that focusing focusing or convergence of eyes is abnormal, causing eye fatigue.

The 3D transformation technique of the 2D image detects the monocular cue from the acquired image, obtains depth information, and converts the binocular cue into binocular cues. When using a general three-dimensional display that exists at a certain depth among binocular cues, convergence of the gaze cannot be used, and transformation using binocular disparity is common.

Various algorithms have been proposed, such as transform using delay time and transform using depth calculation.

Such a conventional algorithm has a problem in that proper three-dimensional transformation is not performed by unilaterally performing three-dimensional transformation without considering the movement of the camera, in particular, the operation such as zooming in and zooming out of the camera. In particular, in the conventional algorithm, there is a problem in that the movement of the camera cannot be converted into an accurate 3D image.

In order to solve the problems of the prior art as described above, an apparatus and method for converting a 2D image into a 3D image in consideration of the movement of the camera is proposed.

Another object of the present invention is to propose an apparatus and method for converting a 2D image into a 3D image using scene structure estimation when the camera moves.

According to an aspect of the present invention, a feature point extraction unit for dividing a region of a 2D image and extracting feature points of the divided region; A camera motion recognition unit for recognizing camera motion in the 2D image; A motion parallax converter configured to generate a disparity image of the 2D image through motion parallax if the camera motion is not recognized by the camera motion recognition unit; A scene structure estimation converter configured to generate a disparity image of the 2D image through scene structure estimation when the camera motion is recognized by the camera motion recognition unit; And an image synthesizer configured to synthesize a 3D image by using the disparity image generated by the motion parallax transform unit or the scene structure estimation transform unit.

The feature point extracting unit may include: a color dividing unit dividing an area of the 2D image by using color information; A labeling unit which assigns a label number to pixels connected in the image divided by the color dividing unit; An outline extractor configured to extract an outline of the labeled image; And a feature point acquisition unit for selecting a feature point for each of the divided regions using the contour extraction information.

The color dividing unit performs color dividing using an average shift algorithm.

The labeling unit searching for pixels having no label number assigned to the color-divided two-dimensional image; Assigning a new label number to the searched pixel; And assigning the same label number to the pixels connected to the searched pixel.

The feature point acquisition unit randomly selects feature points for each of the divided regions or selects the frequency information using frequency information.

The number of feature points selected by the feature point acquirer for each region is based on the size of the region.

The camera motion recognition unit searches for motion vectors in the image boundary regions, and determines that the camera motion panning operation is performed when the searched motion vectors have a constant direction in at least three boundary regions.

The camera motion recognizing unit searches for a motion vector at a corner of the image boundary region, and determines that the camera motion in or out is performed when the searched motion vector has a direction inward or outward.

The camera motion recognizing unit obtains a motion size from each pixel of the border area when searching for a motion vector in the border area, and sets the largest motion size as a motion representative value of the border area, and the representative value is greater than or equal to a preset threshold. I think there is movement.

The motion parallax three-dimensional converter tracks the movement of the feature point using a Kande-Lucas-Tomasi (KLT) feature point tracker.

The motion parallax converter estimates motion by calculating a feature point position difference between frames through a feature point tracking result using the Kande-Lucas-Tomasi (KLT) feature point tracker.

The motion parallax converter generates a disparity image by converting the motion estimation information into disparity.

The scene structure estimation converter calculates depth information of the 2D image by setting a camera model using camera internal parameters including a focal length of the camera and calculating a state vector using an extended Kalman filter.

According to another aspect of the present invention, the method comprises: dividing a region of a 2D image and extracting feature points of the divided region; Recognizing camera movement in the two-dimensional image (b); (C) generating a disparity image for the 2D image through motion parallax if the camera movement is not recognized in the step (b); Generating a disparity image of the 2D image by estimating a scene structure when the camera movement is recognized in the step (b); And synthesizing a three-dimensional image using the disparity image generated in the step (c) or the step (d).

Hereinafter, with reference to the accompanying drawings will be described in detail a preferred embodiment of the method and apparatus for three-dimensional conversion of a two-dimensional image using motion parallax and scene structure estimation according to the present invention.

1 is a block diagram illustrating a configuration of a 3D conversion apparatus of a 2D image according to an exemplary embodiment of the present invention.

Referring to FIG. 1, a 3D transformation apparatus of a 2D image according to an embodiment of the present invention may include a feature point extractor 100, a camera motion recognition unit 102, a motion parallax 3D converter 104, and a scene structure. The estimation 3D converter 106, the hole filling unit 108, and the image synthesizing unit 110 may be included.

The feature point extractor 100 extracts a feature point from a two-dimensional image to be converted (hereinafter, referred to as a 'conversion target image'). The present invention proposes a method for converting a 2D image into a 3D image using motion parallax or scene structure estimation, and the feature point extraction unit 100 needs to extract feature points as a preprocessing operation for estimating the motion parallax and scene structure. ) Performs this function.

The feature point extractor 100 splits the image to be converted and extracts the feature points in units of the divided regions. According to a preferred embodiment of the present invention, a color to be converted is divided into a plurality of regions by using color segmentation.

The feature point extractor 100 divides the region of the conversion target image for each color and performs labeling on the divided region. Through labeling, the divided regions of the image to be converted are sorted in size order.

The feature point extractor 100 obtains the contour of each region after labeling the divided regions and extracts the feature points for each region. According to an embodiment of the present invention, feature points may be randomly extracted for each region. According to another embodiment of the present invention, the feature point may be selected from points having a high frequency component in each region. The number of feature points selected for each region is one or more, and may be determined according to the size of the region. More detailed feature extraction algorithm will be described in detail with a separate drawing.

The camera motion recognition unit 102 distinguishes between the stop and the movement of the camera through the motion analysis of the converted target image. That is, the camera motion recognition unit 102 determines whether the camera moves, such as panning, zooming in, and zooming out. A more detailed method of recognizing the movement of the camera will be described in detail through separate drawings.

When there is camera movement, the camera motion recognition unit 102 causes the image change operation to be performed by the scene structure estimation transformation unit 106. When there is no camera movement, the camera motion recognition unit 102 transmits the motion parallax to the three-dimensional conversion unit 106. This allows image conversion to be performed.

The motion parallax three-dimensional transform unit 104 generates a disparity image of the transform target image by estimating the motion of a feature point in the transform target image and assigning the disparity to the estimated motion. The motion parallax three-dimensional converter 104 estimates the motion based on the area divided by the feature point extractor 100.

The motion parallax three-dimensional converter 104 determines the magnitude of the horizontal motion and the vertical motion of the transform target image during motion estimation. According to an embodiment of the present invention, the variation of the vertical component cannot obtain a sense of depth. Therefore, the motion size in the horizontal direction and the motion size in the vertical direction are normalized and converted into horizontal variations.

The scene structure estimation three-dimensional transform unit 106 obtains the scene structure from the transformation target image to obtain depth information of the projection to generate a disparity image. According to a preferred embodiment of the present invention, the scene structure estimation three-dimensional transform unit 106 obtains the depth of projection by measuring the scene structure and the camera movement using the extended Kalman filter. A more detailed method of acquiring the depth of an image using scene structure estimation will be described in detail with reference to separate drawings.

The hole filling unit 108 may fill a hole generated in the disparity image generated by the motion parallax three-dimensional transform unit 104 and the scene structure estimation three-dimensional transform unit.

When generating a disparity image using motion parallax or scene structure estimation, depth information may not be grasped in all regions divided by the feature point extractor 100, and in this case, holes may occur. As a cause of the occurrence of a hole, movement other than the structure in which motion estimation of a feature point included in the divided region fails or stops may occur because it is not used for scene structure estimation.

When the hole is generated, the hole filling unit 108 compares the color of the area where the hole is generated and the area around the hole to perform hole filling through color interpolation.

The image synthesizer 110 generates a stereo image by synthesizing the disparity images generated by the motion parallax 3D transformer 104 and the scene structure estimation 3D transformer 106 and interpolated by the hole filling unit 108. Function.

2 is a block diagram showing the configuration of a feature point extraction unit according to an embodiment of the present invention.

Referring to FIG. 2, the feature point extractor 200 divides an area of the transform target image by using color information of the transform target image.

According to a preferred embodiment of the present invention, the color dividing unit 200 performs color dividing by using a mean shift algorithm. The average shift algorithm is an algorithm that can measure the density change, and has a relatively simple characteristic since it does not require a parameter.

The average shift algorithm is used to find the midpoint of a region with high density in a specific space, and a sphere-shaped window with a single diameter is defined to minimize the parameters of the search region. This characteristic is in contrast to the conventional algorithm. The conventional algorithm requires the size and shape of the kernel or the number of neighboring pixels to be included in the search, but the average shift algorithm minimizes the input of additional parameters.

의 확률 분포 함수

를 가정한다. 반지름이

, 중점이

인 구

가 특징 벡터

를 포함할 경우

으로 나타낼 수 있다.

와

가 주어졌을 경우

의 기대값은 다음의 수학식 1과 같다. The principle of the mean shift algorithm is as follows. First feature vector

Probability distribution function

Assume Radius

, Emphasis

population

Autumn feature vector

If you include

It can be represented as

Wow

Is given

Expected value is equal to the following Equation 1.

가 충분히 작을 경우, 수학식 1은 다음의 수학식 2와 같이 근사화 될 수 있다.

Is small enough, Equation 1 may be approximated as Equation 2 below.

는 구의 부피를 나타내며,

의 근사값은 다음의 수학식 3과 같다.

Represents the volume of the sphere,

An approximation of is given by Equation 3 below.

는

의 확률 분포 함수의 그래디언트이므로 다음의 수학식 4를 얻는다.

Is

Since it is a gradient of the probability distribution function of, the following equation (4) is obtained.

By integrating Equation 4 above, the following Equation 5 can be obtained.

의 확률 분포의 그래디언트에 비례하는 것을 알 수 있다. 높은 밀도의 영역에서는 큰 값의

와 작은 값의

를 가지게 되므로 적은 평균 이동이 일어나게 되고 수렴점을 찾을 수 있다. 평균 이동 알고리듬의 적용은 먼저 탐색 윈도우의 반지름

과 위치를 정한 후, 평균 이동 벡터를 구하고 그 값만큼 탐색 영역을 이동하여 벡터가 수렴할 때까지 반복적으로 수행한다. 도 3은 본 발명의 일 실시예에 따른 색상 분할이 완료된 영상을 도시한 도면이다. 도 3에서(a)는 색상 분할 전의 변환 대상 영상을 도시한 것이며, (b)는 색상 분할 후의 변환 대상 영상을 도시한 것이다. In other words, the mean moving vector, which is the difference between the regional mean and the midpoint of the search region,

It can be seen that it is proportional to the gradient of the probability distribution of. In areas of high density,

With a small value

Since we have a small average shift, we can find the convergence point. The application of the mean shift algorithm is the radius of the navigation window.

After determining the and positions, the average motion vector is obtained, and the search area is moved by the value, and iteratively performed until the vector converges. 3 is a diagram illustrating an image in which color separation is completed according to an embodiment of the present invention. In FIG. 3, (a) shows a conversion target image before color division, and (b) shows a conversion target image after color division.

When color division is performed by the color dividing unit 200, the labeling unit 202 performs a labeling operation of assigning the same number to all the pixels connected after the color dividing.

4 is a diagram illustrating an example of performing labeling on a target image to be converted according to an exemplary embodiment of the present invention. Labeling may be performed through the following process.

i) When a pixel P (see FIG. 4) that is not yet labeled through image search is found, a new label number is assigned to the pixel.

ii) The same label number is assigned to the pixel connected to the pixel P as shown in FIG.

iii) The process of i) and ii) is repeated to give the same label number to all the pixels connected to the lighted pixel to be labeled.

iv) Repeating steps ii) and iii) to continue labeling until there are no unlabeled pixels.

v) Return to the process of i) to search for the pixels not yet labeled and repeat the process of ii) and iv).

5 is a diagram illustrating an image after labeling the image of FIG. 3 according to an embodiment of the present invention.

The contour extracting unit 204 performs an operation of acquiring the contour in order to select a feature point for the labeled image. The outline extractor 204 calculates the number of pixels included in the labeled image, arranges the order of the labeled objects according to the number of pixels, and extracts the outline.

FIG. 6 is a diagram illustrating an example of finding a boundary for contour extraction according to an embodiment of the present invention. Referring to FIG. 6, the process of extracting the contour is as follows.

i) The image is searched as shown in FIG. 6 to search the boundary point a ₀ without adding the trace completion mark (255 in FIG. 6).

ii) a ₀ if the surroundings are propriety black (0) and point a0 is isolated and finished tracing.

iii) In other cases, the following boundary point is found and the boundary point is traced by the above-described method.

iv) Complete the trace when the next boundary point is a ₀ .

FIG. 7 illustrates an image of completing contour extraction on the labeled image of FIG. 5.

The feature point extractor 206 selects a feature point from the image from which the contour is extracted. As described above, the feature point can be extracted by a method of randomly selecting the feature point or selecting a high frequency component for each divided region. FIG. 8 is a diagram illustrating an image from which feature points are extracted from the image of FIG. 7.

Regarding camera motion recognition, if the same method is applied to all images without considering the camera movement in the 3D transformation method of the 2D image, the object and the background are separated for the image with the camera movement such as a panning camera. it's difficult. This is because the movement of the object is relatively small compared to the movement of the background. In this case, if the fixed camera method is applied in the same way, the three-dimensional effect cannot be felt in the object part. Therefore, a new three-dimensional transformation technique should be applied to the image with camera movement. This requires a process of recognizing the camera movement. The present invention proposes an algorithm that can recognize camera panning and zooming.

Motion vectors are used to recognize camera movements such as still, pan, and zoom from images. The motion analysis method includes a simple method using a difference between two images and an optical flow method using a change in a motion that occurs during a certain time. The optical flow method includes a method of estimating model parameters and a method of analyzing the optical flow pattern from the distribution of the magnitude or angle of the optical flow vector without using a motion model. However, this method is computationally expensive and requires several thresholds. Alternatively, the MPEG video stream is used to recognize camera movement using motion vectors associated with P and B frames. However, this method generally has a limitation on the size of a moving object.

9 is a flowchart illustrating a process of recognizing whether there is a camera panning in a camera motion recognition unit according to an exemplary embodiment of the present invention.

Referring to FIG. 9, a maximum motion vector of a boundary region of a transformation target image is searched (operation 900). In the present invention, in order to recognize the motion of the camera, first, the magnitude of the motion is obtained for each pixel of the boundary portion of the image. In addition, the magnitude of the most frequently generated motion is designated as the motion representative value of the boundary region. If the representative value is greater than or equal to the threshold, it is determined that camera movement has occurred.

Since the method of the present invention obtains motion vector information on the boundary of the image, the calculation amount is smaller than that of the conventional method.

It is determined whether the searched motion vector has a constant direction (step 902). According to a preferred embodiment of the present invention, it is determined whether the motion vectors of the boundary regions of the remaining three portions except the lower boundary region of the image have a constant direction.

If the motion vector is of constant orientation, it is determined that there is camera panning (step 904). If the motion vector does not have a constant direction, it is determined that the motion vector is a fixed camera rather than a camera panning (step 906).

10 is a flowchart illustrating a process of determining camera zoom-in and camera zoom-out according to an exemplary embodiment of the present invention.

Referring to FIG. 10, a motion vector is searched for in a corner region of a boundary region of an image (step 1000). While camera panning searches for motion vectors of all boundary regions, it is preferable to search for motion vectors in corner regions of the boundary regions when determining camera zoom-in and zoom-out.

It is determined whether or not directionality is detected (motion vector is greater than or equal to a preset threshold) at three or more of four corners (step 1002).

If the orientation is not detected at three or more corners, it is determined that the camera is a fixed camera in which zooming in and zooming out do not occur (step 1004).

When directionality is detected at three or more corners, it is determined whether there is a direction toward the inside of the image or a direction toward the outside of the image (step 1006).

If there is directionality inside the image, it is determined that a camera zoom out has occurred (step 1008).

If there is directivity out of the image, it is determined that camera zoom-in has occurred (step 1010).

11 is a flowchart illustrating a process of generating a disparity image using motion parallax according to an exemplary embodiment of the present invention.

Referring to FIG. 11, a feature point tracking procedure is first performed (step 1100). Tracking based on feature points refers to a process of selecting an area to be tracked in an image and finding a corresponding point in each frame of the video.

According to a preferred embodiment of the present invention, feature points are tracked using a Kanade-Lucas-Tomasi (KLT) feature point tracker.

The motion information is obtained and a 3D image is generated by converting the obtained motion vector into a disparity vector.

, 움직임 벡터를

, w(x)=1이라고 할 때, 움직임 측정 오차인

은 다음의 수학식 6과 같이 정의된다. The principle of feature point tracking using KLT is as follows. Assume two images between adjacent frames as I and J, and coordinate the feature points

Motion vector

, w (x) = 1, the motion measurement error

Is defined as in Equation 6 below.

와

를 가정하면 다음의 수학식 7과 같이 정의된다. By primary Taylor watering,

Wow

Suppose that is defined as Equation 7 below.

The derivative is performed as shown in Equation 8 to find a motion vector that minimizes the motion estimation error.

After summarizing the above Equation 8, the following equation 9 finds d that minimizes the motion estimation error.

As described above, when the motion vector information at the feature point is obtained, the motion is estimated by calculating the position difference between the feature points between the frames (step 1102). When estimating the motion, the horizontal motion and the horizontal motion are obtained.

According to a preferred embodiment of the present invention, feature point tracking is performed between 5 and 20 frames depending on the characteristics of the image, and accurate results are obtained by using bidirectional tracking. In other words, the feature point selected in the first frame is traced to obtain the feature point in the last frame, and the feature point trace result is used when the feature point is tracked in the reverse order of the frame and has a similar position within the error range.

When the motion estimation is complete, the motion estimation information is transformed to disparity (step 1106). As described above, according to the preferred embodiment of the present invention, since a sense of depth cannot be obtained from the variation of the vertical component, the vertical movement magnitude and the horizontal movement magnitude are normalized and converted into horizontal displacement.

When the motion information is converted into disparity information, a disparity image is generated using the disparity (step 1106).

12 is a flowchart illustrating a process of generating a disparity image using scene structure estimation according to an exemplary embodiment of the present invention.

Referring to FIG. 12, a camera model for scene structure estimation is set (step 1200).

는 초점 거리의 역수이다. 13 is a diagram illustrating a camera model for scene structure estimation, according to an embodiment of the present invention. When the position in the camera coordinate system is (Xc, Yc, Zc), and the coordinate of the feature point is (u, v), the following equation (10) holds. Camera model according to an embodiment of the present invention uses the camera focal length as a parameter,

Is the inverse of the focal length.

는 스케일 파라미터이다. The scene structure model in the three-dimensional coordinate system (X, Y, Z) can be expressed as Equation 11 below, where

Is a scale parameter.

For scene structure estimation, using F frames and N feature points, 6 (F-1) motion parameters and 3N structure parameters are required. Since 2NF information and one scale information are needed according to the number of feature points, the condition of 2NF + 1> 6 (F-1) + 3N must be satisfied to solve the problem. Therefore, the conditions of F ≥ 2 and N ≥ 6 must be satisfied to solve the motion and scene structure estimation. On the other hand, the extended Kalman filter does not take F into account because the solution is solved through a recursive solution. According to the number of feature points, 2N information and 1 scale information are required, and 6 camera external parameters, 1 focal length, and N structural parameters are required, so that a solution can be obtained when N> 7.

Once the camera model is set up, the relationship between the coordinate system and the three-dimensional coordinate system in the camera model is set (step 1202).

The relationship between the camera coordinate system and the three-dimensional coordinate system may be set as in Equation 12 below.

The relationship between the three-dimensional coordinate system and the camera coordinate system is defined as in Equation 12 (1), which is a parallel movement and rotational motion. The rotational motion is defined as in Equation 12 (2), and the rotational motion is defined as in Equation 12 using Equation 4 (3).

Six camera external parameters, one camera internal parameter, and N scene structures are configured by a state vector as shown in Equation 13.

가 포함되지 않은 결과이며, 깊이 정보의 평균고 표준 편차를 이용하여 0에서 255 사이의 값을 가지는 스케일링 파라미터

를 구한다(단계 1206). The extended Kalman filter can be used to obtain the state vector X and the measurement vector, and can also acquire depth information at the feature points of the scene structure (step 1204). However, this depth information is not a scaling parameter.

Is a result that does not contain a scaling parameter with a value between 0 and 255 using the average high standard deviation of the depth information.

Is obtained (step 1206).

는 다음의 수학식 14에 구해질 수 있다. Scaling parameters with values between 0 and 255

Can be obtained from Equation 14 below.

는 장면 구조 값의 표준 편차이고

는 장면 구조 값의 평균이다. In Equation 14 above

Is the standard deviation of the scene structure values

Is the average of the scene structure values.

를 이용하여 장면 구조 추정에 의한 변이 영상을 획득한다. Scaling parameters

Acquire a disparity image by scene structure estimation using.

As described above, according to the preferred embodiment of the present invention, an adaptive three-dimensional image conversion is possible by converting a two-dimensional image into a three-dimensional image in consideration of the movement of the camera.

In addition, according to a preferred embodiment of the present invention, by using the long-life structure estimation for the three-dimensional transformation when the camera movement, there is an advantage that the more efficient three-dimensional image conversion by using the scene structure estimation when the camera movement .

Claims (26)

A feature point extractor for dividing a region of the 2D image and extracting feature points of the divided region; A camera motion recognition unit for recognizing camera motion in the 2D image; A motion parallax converter configured to generate a disparity image of the 2D image through motion parallax if the camera motion is not recognized by the camera motion recognition unit; A scene structure estimation converter configured to generate a disparity image of the 2D image through scene structure estimation when the camera motion is recognized by the camera motion recognition unit; And And an image synthesizer configured to synthesize a 3D image using the disparity image generated by the motion parallax transform unit or the scene structure estimation transform unit. The method of claim 1, The feature point extraction unit, A color dividing unit dividing an area of the 2D image using color information; A labeling unit which assigns a label number to pixels connected in the image divided by the color dividing unit; An outline extractor configured to extract an outline of the labeled image; And And a feature point acquisition unit for selecting a feature point for each of the divided regions by using the contour extraction information. 2. The method of claim 2, And the color dividing unit performs color dividing using an average shift algorithm. The method of claim 2, The labeling unit, Searching for pixels having no label number assigned to the color-divided two-dimensional image; Assigning a new label number to the searched pixel; And And assigning the same label number to the pixels connected to the searched pixel. The method of claim 2, And the feature point obtaining unit randomly selects the feature points for each of the divided regions or uses frequency information to select the feature points. The method of claim 5, And the number of feature points selected by the feature point acquirer for each region is based on the size of the region. The method of claim 1, The camera motion recognition unit searches for a motion vector in the image boundary regions, and if the searched motion vector has a constant direction in at least three boundary regions, the camera motion recognition unit determines that the camera panning operation is performed. Video conversion device. The method of claim 1, The camera motion recognition unit searches for a motion vector at a corner of an image boundary region, and if the searched motion vector has a direction inward or outward, it is determined as a camera zoom-in or a camera zoom-out operation. 3D image conversion device. The method according to claim 7 or 8, The camera motion recognizing unit obtains a motion size from each pixel of the border area when searching for a motion vector in the border area, and sets the largest motion size as a motion representative value of the border area, and the representative value is greater than or equal to a preset threshold. A three-dimensional image conversion apparatus for a two-dimensional image, characterized in that the movement is determined. The method of claim 1, And the motion parallax three-dimensional converter tracks the movement of the feature point using a Kande-Lucas-Tomasi (KLT) feature point tracker. The method of claim 10, The motion parallax transform unit estimates motion by calculating a feature point position difference between frames through a feature point tracking result using the Kande-Lucas-Tomasi (KLT) feature point tracker. The method of claim 11, And the motion parallax converter converts the motion estimation information into a disparity to generate a disparity image. The method of claim 1, The scene structure estimation converter calculates depth information of the 2D image by setting a camera model using camera internal parameters including a focal length of the camera and calculating a state vector using an extended Kalman filter. 3D image conversion device of 3D image. The method of claim 13, The depth information is represented by the following equation

Is computed by

Is the standard deviation of the scene structure values,

Is a mean of scene structure values. Dividing a region of the 2D image and extracting feature points of the divided region; Recognizing camera movement in the two-dimensional image (b); (C) generating a disparity image for the 2D image through motion parallax if the camera movement is not recognized in the step (b); Generating a disparity image of the 2D image by estimating a scene structure when the camera movement is recognized in the step (b); And And synthesizing a three-dimensional image using the disparity image generated in the step (c) or the step (d). The method of claim 15, Step (a) is, A color segmentation step of segmenting an area of the 2D image by using color information; A labeling step of assigning label numbers to pixels connected in the image divided by the color division; Contour extraction step of extracting a contour on the labeled image; And And a feature point obtaining step of selecting a feature point for each of the divided regions using the contour extraction information. The method of claim 16, And the color dividing unit performs color dividing using an average shift algorithm. The method of claim 16, The labeling step,  Searching for pixels having no label number assigned to the color-divided two-dimensional image; Assigning a new label number to the searched pixel; And assigning the same label number to the pixels connected to the searched pixel. The method of claim 16, The method for acquiring a feature point may include selecting a feature point randomly for each of the divided regions or by using frequency information. The method of claim 15, The step (b) is to search for motion vectors in the image boundary regions, and if the searched motion vectors have a constant direction in at least three boundary regions, it is determined as a camera panning operation. Dimensional Image Conversion Method. The method of claim 15, The step (b) is to search for a motion vector in the corner portion of the image boundary region, and if the searched motion vector has a direction inward or outward, it is determined that the camera zoom in or camera zoom out operation 3D image conversion method. The method of claim 15, The step (c) is a three-dimensional image conversion method of a two-dimensional image, characterized in that to track the movement of the feature point using a Kande-Lucas-Tomasi (KLT) feature point tracker. The method of claim 22, The step (c) is a three-dimensional image transformation method of a two-dimensional image, characterized in that the motion is estimated by calculating the position difference between the feature points between the frame using the feature point tracking results using the Kande-Lucas-Tomasi (KLT) feature point tracker. The method of claim 22, The step (c) is a three-dimensional image conversion apparatus for a two-dimensional image, characterized in that for generating a disparity image by converting the motion estimation information into a disparity. The method of claim 1, In step (d), the depth information of the 2D image may be calculated by setting a camera model using camera internal parameters including a focal length of the camera and calculating a state vector using an extended Kalman filter. 3D image conversion method of 2D image. The method of claim 25, The depth information is represented by the following equation

Is computed by

Is the standard deviation of the scene structure values,

Is a mean of scene structure values.

Images