KR101667306B1 - Low-Complexity Topological Derivative-Based Image Segmentation Method and System - Google Patents

Abstract

Provided are TD-based low-complexity image segmentation method and a system thereof. An image segmentation method according to the present invention calculates cost by putting an original image and an updated segment image in a cost function. Filters which show calculators of the cost function are represented with at least one of bilateral and vertical symmetric matrixes. Thereby, high performance and low complexity can be processed with regard to image segmentation such as broadcasting, content production, and medical images.

Description

[0001] The present invention relates to a low-complexity topological derivation-based image segmentation method and system,

The present invention relates to image segmentation, and more particularly, to a method and system for image segmentation based on TD (Topological Derivative).

The cost function used for TD-based image segmentation is expressed by the following equation.

The above cost function is expressed in the digital domain as follows.

As can be seen from the above, the sparse matrix used for the cost function is very complicated and many matrix multiplication (x) operations are required for the operation of the cost function, so that the cost calculation is very complicated . And this is a factor to lower the TD-based image segmentation speed.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a TD-based image segmentation method and system capable of drastically reducing complexity without deteriorating performance.

According to an aspect of the present invention, there is provided an image segmentation method comprising: setting an original image as an initial segmentation image; Updating a segmentation image; Calculating a cost by substituting an original image and an updated segmentation image into a cost function; And determining that the updated segmentation image is a final segmentation image if the calculated cost is less than the set threshold value, and wherein the operators constituting the cost function are represented by filters that are symmetric with respect to at least one of up, down, left, and right .

In the first operator that constitutes the cost function, all coefficients except the coefficient located at the center may be represented by the same filter.

In addition, the second operator that forms the cost function may have the same coefficients as those located at the vertices, and the remaining coefficients except for the coefficients located at the center may be represented by the same filter.

The cost function may include at least one of a convolution (*) operation and an inter-element multiplication (·) operation without including a matrix multiplication operation.

Further, the cost function is expressed by the following equation,

K _H and M _H may be the filters.

And, K _H and M _H can be expressed by the following filters.

Also, a ₁ , C ₁ , C ₂ , and C ₃ may be positive coefficients, and a ₂ may be a negative coefficient.

According to another aspect of the present invention, an image system includes an input unit for receiving an original image; And setting the original image input through the input unit as an initial segmentation image, updating a segmentation image, substituting an original image and an updated segmentation image into a cost function to calculate a cost, and if the calculated cost is less than a preset threshold And a processor for determining an updated segmentation image as a final segmentation image, wherein the operators constituting the cost function are represented by filters whose at least one of up, down, left and right is symmetrical.

As described above, according to the embodiments of the present invention, image segmentation used for broadcasting, content production, medical image, and the like can be processed with low complexity with high performance.

Particularly, according to the embodiments of the present invention, there is almost no degradation in performance while drastically reducing the complexity compared to the conventional technique.

In addition, according to the embodiments of the present invention, since a large storage space is not required for the calculation, the memory can be processed in parallel with relatively small GPUs, so that the image segmentation speed can be further improved.

FIG. 1 is a flow chart for explaining an image segmentation method according to an embodiment of the present invention;
FIG. 2 is a block diagram of an image system capable of performing the image segmentation method shown in FIG. 1,
FIG. 3 is a diagram showing the procedures performed by the GPU and the procedures performed by the CPU,
4 is a diagram showing a method of processing an original image of a CPU and a GPU,
5 is a diagram illustrating a pseudo code used for cost calculation in a GPU,
6 is a diagram showing a segmentation performance result,
7 to 9 are tables showing the performance of the segmentation method according to the embodiment of the present invention.

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

In the embodiment of the present invention, an image segmentation method based on TD (Topological Derivative) can dramatically reduce the complexity compared to the conventional method.

Specifically, in the image segmentation method according to the embodiment of the present invention, the sparse matrices used in the cost function of the existing technique are replaced with the 2D filters of the small size, and the matrix multiplication (x) Operations are replaced by convolution operations and element by element multiplication operations.

FIG. 1 is a flowchart illustrating an image segmentation method according to an exemplary embodiment of the present invention. Referring to FIG.

As shown in FIG. 1, when the original image v is input, the original image v is set as an initial value of the segmentation image u to be calculated (S110).

)를 계산한다(S120). VPE는 다음의 수학식으로 표현된다.Then, the solution of the VPE (Variational Problem Equation)

(S120). VPE is expressed by the following equation.

Here, K is a Diffusion Tensor (DT), and? Is a weight parameter between 0 and 1. eta is removed as a weakening.

Then, TD (Topological Derivative: D _T ) is calculated (S130). TD (D _T ) is expressed by the following equation.

Here, C _i represents the kinds of segments appearing in the segmentation image u.

Next, the sorting by the TD (D _T) and using (S140), converts the result of sorting some of the pixel values of the segmentation image (u), updating the segmentation image (u) (S150).

Specifically, in step S140, pixels having a negative (-) TD (D _T ) smaller than the threshold value in the segmentation image u are sorted, and the sorted pixels are changed to different values in step S150.

), 원본 영상(v) 및 세그먼테이션 영상(u)을 코스트 함수에 대입하여 코스트를 계산한다(S160). 코스트 함수는 다음의 수학식으로 표현된다.Then, the solution of VPE (

), The original image (v), and the segmentation image (u) to the cost function to calculate the cost (S160). The cost function is expressed by the following equation.

The above cost function is expressed in the digital domain as follows.

)이고, M_H는 코스트 함수의 두 번째 항인 불일치도(Mismatch Evaluation)를 나타내는 성분의 연산기(

)이다.Here, K _H is an operator of the component representing the smoothness evaluation, which is the first term of the cost function (

) And M _H is the operator of the component representing the mismatch evaluation, which is the second term of the cost function (

)to be.

These are implemented with the following 2D filters.

As shown above, K _H is a 3 × 3 filter in which the coefficients constituting the filter are vertically symmetric (symmetrical with respect to two rows), left and right displaced (symmetrical with respect to two columns), diagonal symmetry (symmetrical with respect to diagonal) to be. Where a ₁ is a positive filter coefficient, and a ₂ is a negative filter coefficient. K _H is a filter having the properties of a differentiator. K _H are all the filter coefficients except for the filter coefficients located at the center of the matrix.

Similarly to K _H , the coefficients constituting the M _H filter are divided into 3 × 3 filters (symmetric with respect to the diagonal), left and right substitution (symmetrical with respect to the second row), and diagonal symmetry to be. C ₁ , C ₂ , and C ₃ are both positive filter coefficients. M _H is a filter with the properties of an integrator. M _H is the vertex part [(1, 1), 1.3), (3.1), (3.3) filter coefficients are the same, the upper part of the filter coefficient which is located at the center (1,2) which is located in the lower (3, 2), the filter coefficients located on the left (2,1) and right (2,3) are the same.

In addition, it can be predicted that the above cost function will be significantly reduced in that the above cost function does not include a matrix multiplication (×) operation but is expressed by a convolution (*) operation and an inter-element multiplication (·) operation.

If the cost calculated in step S160 is less than the set threshold (S170-Y), the updated segmentation image u is determined as the final segmentation image in step S150 (step S180).

However, if the cost calculated in step S160 is equal to or greater than the preset threshold value (S170-N), the process is resumed from step S120 to find a proper segmentation image u.

2 is a block diagram of an image system capable of performing the image segmentation method shown in FIG. As shown in FIG. 2, the image system includes an image input unit 110, a processor 120, and an image output unit 150.

The video input unit 110 provides the original video v to the processor 120. [ The video input unit 110 is a communication interface for receiving an original video v from a storage medium storing an original video v or an external device or network.

The processor 120 performs the procedure of FIG. 1 for segmenting the original image v input from the image input unit 110. FIG. The image output unit 150 is a means for outputting the segmentation image applied from the processor 120 through a display or transmitting the segmentation image to an external device or a network.

The processor 120 includes a GPU 130 and a CPU 140. In FIG. 3, the procedures performed by the GPU 130 and the procedures performed by the CPU 140 among the procedures shown in FIG. 1 are separately shown.

4, the GPU 130 divides the original image and performs parallel processing, unlike the CPU 140. [ In addition, FIG. 5 shows a pseudo code used for the cost calculation in the GPU 130.

6 is a diagram showing a segmentation performance result. 6 (a) and 6 (e) are original MRI images, (b) and (f) are segmentation results using a conventional sparse matrix, and (c) and (g) (D) and (h) are the results of performing the segmentation according to the embodiment of the present invention by CPU + GPU (i.e., parallel processing).

As shown in the table shown in FIG. 7, it can be seen that the complexity of the segmentation method according to the embodiment of the present invention is remarkably reduced as compared with the complexity of the segmentation method according to the existing method.

Also, as shown in the table shown in FIG. 8, the segmentation method according to the embodiment of the present invention shows no significant difference in the segmentation method and the segmentation result according to the existing method.

Further, as shown in the table shown in FIG. 9, it can be seen that the computation speed is much faster to perform with CPU + GPU (i.e., parallel processing) than to perform segmentation with CPU alone Can be confirmed.

Up to now, a TD-based low complexity image segmentation method has been described in detail with a preferred embodiment.

Although it has been assumed in the above embodiment that the medical image is segmented, it goes without saying that the technical idea of the present invention can also be applied to other images (for example, a broadcast image and a camera image).

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, but, on the contrary, It will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the present invention.

110:
120: Processor
130: GPU
140: CPU
150: Video output unit

Claims (8)

Setting an original segmentation image as an initial segmentation image;
Updating a segmentation image;
Calculating a cost by substituting an original image and an updated segmentation image into a cost function; And
Determining an updated segmentation image as a final segmentation image if the calculated cost is less than the set threshold,
The arithmetic operators constituting the cost function,
Wherein at least one of up, down, left and right is represented by symmetric filters.
The method according to claim 1,
The first operator, which constitutes the cost function,
Wherein all coefficients except for the coefficients located at the center are expressed by the same filter.
The method according to claim 1,
The second operator, which constitutes the cost function,
The coefficients located at the vertices are the same,
Wherein coefficients except for the coefficients located at the center or the vertex are expressed by the same filter.
The method according to claim 1,
The cost function includes:
Without including matrix multiplication,
A convolution (*) operation and an inter-element multiplication (·) operation.
The method of claim 4,
The cost function is expressed by the following equation,

ψ '(Ω) is the cost function,

Is the solution of VPE (Variational Problem Equation)
v is the original video,
u is a segmentation image,
K _H and M _H are said filters,
And N x M is a filter size.
The method of claim 5,
K _H and M _H are expressed by the following filters.

The method of claim 6,
a ₁ , C ₁ , C ₂ , C ₃ are positive coefficients,
and a ₂ is a negative coefficient.
An input unit for receiving an original image; And
The method includes setting the original image input through the input unit as an initial segmentation image, updating a segmentation image, substituting an original image and an updated segmentation image into a cost function to calculate a cost, and if the calculated cost is less than a set threshold, And determining the segmentation image as a final segmentation image,
The arithmetic operators constituting the cost function,
Wherein at least one of up, down, left and right is represented by symmetrical filters.

Images