KR101691380B1 - Dct based subpixel accuracy motion estimation utilizing shifting matrix - Google Patents

Abstract

A motion prediction method for a 1/4 sub-pixel for video encoding, the method comprising: performing DCT (Discrete Cosine Transform) transformation on an input image frame; Performing motion prediction in a frequency domain to obtain a prediction block having a minimum matching error with a current block through block matching between a DCT transformed current image frame and a previous image frame, The prediction block may include a horizontal / vertical shifting matrix applied with a 1/4 sub-pixel interpolation using a three-dimensional interpolation function, a search range of the previous image frame, Is obtained through a matrix operation using a DCT coefficient matrix of four blocks in the vicinity of the prediction block.

Description

[0001] The present invention relates to a DCT-based subpixel motion prediction method using a shifting matrix,

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a moving image compression technique, and more particularly, to a motion estimation method based on a DCT (Discrete Cosine Transform) based sub-pixel using a shifting matrix.

With the recent development of mobile communication terminals, users are expected to enjoy high quality video service even in a wireless environment. Since the capacity for transmitting and receiving a high-quality moving picture in a limited bandwidth communication environment is very large, moving picture compression is indispensably used in order to satisfy a user's requirement. H.264 / AVC, a video compression standard jointly proposed by ITU-T and ISO, is currently the most widely used codec and has been developed to obtain better image quality at a bit rate of less than half the previous video compression standard. As a result, H.264 technology is widely used not only in desktops and laptops but also in digital cameras, mobile phones, and CCTV.

In video compression, techniques for eliminating spatial redundancy and temporal redundancy are used. Among them, motion prediction technology is used as a method for eliminating temporal redundancy between successive frames in an image. Since the motion estimation occupies 60 ~ 80% of the total calculation amount in the video encoder, it is continuously studied to develop a motion prediction algorithm having excellent compression efficiency and low calculation amount. The block matching algorithm, which is the basis of motion prediction, is a method of finding a block having a minimum matching error in a search range of each block of a current frame in a previous frame to predict motion information . The moving picture compression system entropy-codes only the motion information and the error of the prediction block, and transmits the motion vector information as a bitstream. As the motion information is accurate, the error of the prediction block is reduced, thereby effectively eliminating temporal redundancy.

Since the motion information is not generated only in the unit of an integer pixel due to the brightness change between frames, nonlinear motion, boundary between objects, accurate motion information can not be expected by motion prediction of an integer pixel unit only. Since subpixel unit motion prediction can predict subpixel motion information compared to integer pixel unit motion prediction, more accurate motion prediction is possible. For this reason, MPEG-1, MPEG-2 and H.263 use half-pixel motion prediction, and H.264 extends the resolution to 1/4 pixel units to improve motion prediction performance over existing video coding standards.

A motion prediction algorithm of a general video encoder is performed in a spatial domain. However, in the case of subpixel motion prediction in the spatial domain, the application of the interpolation technique increases the total amount of computation of the video encoder.
The prior art related to cubic interpolation in the context of the present invention is "adaptive cubic interpolation considering neighboring pixel values" (The Journal of Broadcasting Engineering, Vol. 15, No. 3, , Lee, Young, Kim, and Jeong, Chang).

The present invention provides a method of performing Motion Estimation of sub-pixel precision based on DCT (Discrete Cosine Transform) using a shifting matrix in a frequency domain.

According to an aspect of the present invention, there is provided a motion prediction method for a 1/4 sub-pixel for video encoding, the method comprising: performing DCT (Discrete Cosine Transform) conversion on an input video frame; Performing motion prediction in a frequency domain to acquire a prediction block having a minimum matching error with a current block through block matching between a DCT transformed current image frame and a previous image frame,

In the step of performing motion prediction in the frequency domain, the prediction block may include a horizontal / vertical shifting matrix to which image interpolation is applied in units of quarter pixels using a three-dimensional interpolation function, And a DCT coefficient matrix of four blocks around the prediction block in a search range of a frame.

In one embodiment, the prediction block is defined by the following equations (1) and (2), and the horizontal shifting matrix uses a matrix defined by Equation (3) and Equation (4), and the vertical shifting matrix A matrix defined by the following equation (5) can be used.

[Equation 1]

는 예측 블록

의 DCT 계수 행렬로 표현됨.
here,

Gt;

Gt; DCT < / RTI >

&Quot; (2) "

의 주변의 4개의 참조 블록 중 어느 하나로서 4×4 블록 사이즈를 가지며, 상기

와

는 각각 상기 수평 시프팅 매트릭스와 상기 수직 시프팅 매트릭스를 나타냄.
Here, f _i is a prediction block

And has a 4x4 block size as any one of the four reference blocks around the reference block,

Wow

Respectively represent the horizontal shifting matrix and the vertical shifting matrix.

&Quot; (3) "

&Quot; (4) "

중 어느 하나의 1/4 부화소 위치를 나타내며,

는 버림 함수를 나타내고,

는 올림 함수를 나타내고, D는

인 4×4의 변위 행렬을 나타냄.
Where f () denotes the three-dimensional interpolation function, d

Pixel sub-pixel position,

Represents a discarding function,

Represents a rounding function, and D represents

4 x 4 displacement matrix.

&Quot; (5) "

대신 수직 변위량인

가 적용됨.
Here, the superscript t represents the transpose matrix, and the discarding function and the rounding function in Equation (5) are the horizontal displacements in Equations (3) and

Instead,

Is applied.

In one embodiment, in the matrix operation for obtaining the prediction block,

의 n 값이 상기 수평 변위량 및 상기 수직 변위량에 따라,

의 값을 갖는 경우

과

은 4×4의 영행렬을 이용하고,

의 값을 갖는 경우

및

의 행렬을 이용할 수 있다.
The variation matrix of Equation (3), Equation (4) and Equation (5)

The n values of the horizontal displacement amount and the vertical displacement amount,

≪ / RTI >

and

A 4 × 4 zero matrix is used,

≪ / RTI >

And

Can be used.

In one embodiment, the three-dimensional interpolation function may be a b-spline interpolation function of Equation (6) or a cubic interpolation function of Equation (7).

&Quot; (6) "

&Quot; (7) "

여기서, 상기 수학식 7의 a는 cubic 보간 함수에 적용될 선택된 상수값임.

In Equation (7), a is a selected constant value to be applied to the cubic interpolation function.

In one embodiment, when DCT coefficient information outside the boundary of the search area is required according to the horizontal displacement amount or the vertical displacement amount, the DCT coefficient information in the search area may be copied and used for the motion prediction.

According to an embodiment of the present invention, a DCT-based motion prediction method having sub-pixel precision using a shifting matrix in a frequency domain is applied to a motion prediction method in a spatial domain It is possible to realize a low calculation amount and a high PSNR (Peak Signal-to-Noise Ratio).

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a diagram showing a structure of a moving picture encoder in a spatial domain; Fig.
2 is a diagram illustrating a structure of a moving image encoder in a frequency domain;
FIGS. 3 and 4 are illustrative drawings for explaining a motion prediction method using a shifting matrix; FIG.
5 is a view illustrating an example of adopting a part of a motion prediction method in an intra prediction mode in a DCT-based subpixel precision motion prediction method using a shifting matrix according to an embodiment of the present invention.
FIGS. 6 and 7 are graphs illustrating performance comparisons between the motion estimation method and the prior art according to an embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS The present invention is capable of various modifications and various embodiments, and specific embodiments are illustrated in the drawings and described in detail in the detailed description. It is to be understood, however, that the invention is not to be limited to the specific embodiments, but includes all modifications, equivalents, and alternatives falling within the spirit and scope of the invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS In the following description of the present invention, detailed description of known related arts will be omitted when it is determined that the gist of the present invention may be unnecessarily obscured. In addition, numerals (e.g., first, second, etc.) used in the description of the present invention are merely an identifier for distinguishing one component from another.

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

In the frequency domain DCT  Based motion prediction

FIG. 1 is a diagram illustrating a structure of a moving picture encoder in a spatial domain, and FIG. 2 is a diagram illustrating a structure of a moving picture encoder in a frequency domain. Hereinafter, advantages of the motion prediction method according to the embodiment of the present invention will be described with reference to FIGS. 1 and 2 by comparing and comparing the motion prediction method in the spatial domain and the motion prediction method in the frequency domain.

Since motion picture coding and decoding are all performed on a block basis, motion estimation is also based on a block-based algorithm. Block - based motion estimation algorithms can be divided into spatial domain algorithms and frequency domain algorithms. The motion estimation algorithm in the spatial domain is composed of a block matching algorithm and a gradient-based algorithm. The motion estimation algorithm in the frequency domain can be composed of a DCT or a wavelet-based motion estimation algorithm. Among the algorithms in the frequency domain, the wavelet-based algorithm consumes less search and matching time. However, since it predicts motion in all subbands, it has a high computational complexity and has to transmit the motion information for all subbands, . On the other hand, the DCT-based algorithm is advantageous in that the existing video compression standard is based on the DCT, and the IDCT process can be omitted because the motion prediction is performed using the DCT coefficients. Accordingly, the embodiment of the present invention employs a DCT-based motion prediction method for performing motion prediction using a DCT coefficient in a frequency domain.

DCT conversion (refer to reference numeral 110 in FIG. 1 and reference numeral 210 in FIG. 2) is a technique for converting information in a spatial domain into a frequency domain. After DCT, energy is concentrated in a low frequency band, It is widely used in bar image compression field. In general, a motion picture coder uses motion prediction in a spatial domain, and FIG. 1 shows a motion picture coder that performs motion prediction in a spatial domain.

Referring to FIG. 1, a moving picture encoder for performing motion prediction in a spatial domain includes a DCT transformer 110, a quantizer 120, an encoder 130 (in this example, an example of an entropy encoder), an inverse quantizer 125), an IDCT transformer (i.e., an inverse DCT transformer) 115, and a motion predictor 140. [ As described above, the moving picture encoding apparatus for performing motion prediction in the spatial domain must store the information of the reconstruction frame used as the reference frame as a spatial region (i.e., convert the frequency-domain-converted information into a spatial region The IDCT process is required.

On the other hand, the motion prediction in the frequency domain uses the DCT transformed DCT coefficients for motion prediction, so that the IDCT transformation process can be omitted. Referring to FIG. 2, the moving picture encoding apparatus for performing motion prediction in the frequency domain includes a DCT transformer 210, a quantizer 220, an encoder 230, an inverse quantizer 225, and a motion estimator 240 , It can be confirmed that the IDCT converter is omitted. Therefore, the motion estimation in the frequency domain is advantageous in that the amount of calculation is reduced due to the structure of the simplified encoder compared with the motion estimation method in the spatial domain.

Shifting  Motion Estimation Using Matrix Algorithm

In the embodiment of the present invention, a shifting matrix is used in the process of performing DCT-based motion prediction in the frequency domain. Hereinafter, a motion prediction method using a shifting matrix will be described with reference to FIGS. 3 and 4. FIG. FIGS. 3 and 4 are reference diagrams for illustrative purposes of the shifting matrix algorithm.

참조)과 가장 유사한 예측 블록(도 3의

참조)을 찾기 위해 그 주변의 4개의 참조 블록(도 3의

참조)이 사용되는 과정을 보여준다. 이와 같이, 시프팅 매트릭스 알고리즘은 특정 예측 블록이 그 주변의 4개의 참조 블록의 수평, 수직 변환된 형태의 합으로 표현할 수 있다는 점을 이용한다. 도 3에서,

및

는 각각 수평 및 수직 방향에서의 변위(픽셀 이동)를 나타내고,

는 움직임 벡터(Motion vector)를 나타낸다. Generally, motion prediction includes a process of finding a prediction frame having a minimum matching error with a specific block of a current frame in a previous frame. FIG. 3 is a block diagram of the current block

(See Fig. 3

(See FIG. 3) in order to find the reference block

) Is used. As such, the shifting matrix algorithm takes advantage of the fact that a particular prediction block can be represented as the sum of the horizontally and vertically transformed forms of the four reference blocks around it. 3,

And

(Pixel shift) in the horizontal and vertical directions, respectively,

Represents a motion vector.

는 수평 시프팅 매트릭스인

와 수직 시프팅 매트릭스인

를 이용하여 하기 수학식 1과 같이 나타낼 수 있다. 여기서, f_i는 예측 블록 주변의 임의 위치의 참조 블록을 나타낸다.
In accordance with this shifting matrix algorithm,

Is a horizontal shifting matrix

And the vertical shifting matrix

Can be expressed by the following equation (1). Here, f _i represents a reference block at an arbitrary position around the prediction block.

와

는 각각 하기 수학식 2 및 수학식 3과 같이 정의될 수 있다. 하기 수학식 2 및 수학식 3에서, 특정 행렬의 위첨자 t는 해당 행렬의 전치행렬(Transposed matrix)을 나타낸다.
At this time,

Wow

Can be defined by the following equations (2) and (3), respectively. In the following equations (2) and (3), a superscript t of a specific matrix represents a transposed matrix of the corresponding matrix.

은 하기 수학식 4와 같이 정의되는 4×4 행렬이다. 이때,

은

의 단위행렬을 나타낸다.
Also here, the displacement matrix

Is a 4x4 matrix defined as: < EMI ID = 3.0 > At this time,

silver

≪ / RTI >

일 경우, 예측 블록

가 예측되는 과정이 도 4를 통해 도시되고 있다. 이때의

와

는 하기 수학식 5 및 수학식 6과 같이 정의될 수 있다.
For example,

, The prediction block

Is predicted as shown in FIG. At this time

Wow

Can be defined by the following equations (5) and (6).

는 DCT 변환행렬을 나타낸다.
At this time, using the orthogonality and separability of the DCT transform, Equation (1) described above can be expressed as Equation (7). here,

Represents a DCT transformation matrix.

의 DCT 계수인

를 구하기 위해서는 8개의 행렬 곱셈과 3개의 행렬 덧셈이 요구된다. 하지만 저주파수 대역에 에너지가 집중되는 DCT 변환 특성을 이용하면 많은 연산량을 감소시킬 수 있다. 따라서 DCT 변환 후 양자화된 참조 블록

의 특성에 기반하여, 계산상의 복잡도를 줄이기 위해 상기 수학식 7은 하기 수학식 8과 같이 유도할 수 있다. 하기 수학식 8은 4×4 블록 크기인 경우를 가정하여, m 및 n이 0 ~ 3까지의 변위값을 갖는 경우로 한정한 것이다. 따라서, 하기 수학식 8에서 m 및 n은 블록 크기에 따라 그 최대 변위값이 변경될 수 있음은 물론이다. 예를 들어, 8×8 블록 크기를 기준으로 움직임 예측이 수행된다면, m 및 n은 모두 0 ~ 7까지의 변위값을 갖게 될 것이다.
According to Equation (7) above,

DCT coefficients of

Requires 8 matrix multiplications and 3 matrix additions. However, using the DCT conversion characteristic in which energy is concentrated in the low frequency band can reduce a large amount of computation. Therefore, the DCT transformed quantized reference block

The above Equation (7) can be derived as Equation (8) below in order to reduce computational complexity. The following equation (8) is limited to a case where m and n have displacement values ranging from 0 to 3, assuming a case of 4 × 4 block size. Therefore, it is a matter of course that m and n in Equation (8) can be changed according to the block size. For example, if motion prediction is performed based on an 8x8 block size, both m and n will have displacement values of 0 to 7.

이 0의 값을 가질 경우 계산에서 제외될 수 있기 때문에, 0이 아닌

계수들의 희소성을 이용하면 예측 블록을 효율적으로 계산할 수 있다. 즉, 일반적인 동영상 압축 알고리즘에서와 마찬가지로 DCT 변환 후 저주파수 대역에 에너지가 집중되므로, 고주파 대역에 있는 계수들은 양자화 이후에는 대부분 0의 값을 갖게 되어 0이 아닌 계수들의 개수는 대폭 줄어들게 된다. 이러한 특징을 이용하기 위해,

계수들(하기 수학식 8의 경우에는, 4×4 블록 크기의 총 4개의 참조 블록을 이용하므로,총 64개의 DCT 계수들이 존재함)을 하기 수학식 9와 같이 지그재그 스캔 순서로 정리할 필요가 있다.
In Equation (8)

Since a value of 0 can be excluded from the calculation,

Using the scarcity of the coefficients, the prediction block can be calculated efficiently. That is, since the energy is concentrated in the low frequency band after the DCT conversion, the coefficients in the high frequency band are mostly 0 after the quantization, so that the number of non-zero coefficients is greatly reduced. To take advantage of this feature,

Coefficients (in the case of Equation (8), a total of 64 DCT coefficients exist because a total of 4 reference blocks having a size of 4x4 blocks are used), it is necessary to sort the DCT coefficients in a zigzag scan order as shown in Equation (9) .

의 개수를 조절하여 효율적인 연산을 할 수 있다. 또한 블록의 크기와

에서 0이 아닌 성분의 개수에 따라 연산량을 효과적으로 줄일 수 있다는 장점을 가진다.
The recursive equation derived from Equation (9)

So that efficient calculation can be performed. Also,

The amount of computation can be effectively reduced according to the number of nonzero components.

Motion prediction with 1/4 pixel accuracy using cubic interpolation function

Generally, in the sub-pixel motion prediction in the spatial domain, motion prediction is performed after the process of enlarging the image to twice or more by using an interpolation technique. In this case, in the case of a half-pixel (i.e., 1/2 pixel) unitary motion prediction using the bilinear interpolation method, the motion prediction performance is low because the interpolation performance is relatively low. On the other hand, in the case of b-spline interpolation or cubic interpolation, a method of predicting a sub-pixel with a cubic function given a specific weight is disadvantageous in that the interpolated value is more accurate than the bilinear interpolation method, . Accordingly, although the H.264 standard can improve the image quality through quarter-pixel motion prediction, the amount of calculation of the motion prediction process is increased due to the application of the interpolation technique, thereby affecting the video coding time. Equation (10) represents a cubic function used in b-spline interpolation, and Equation (11) represents a cubic function used in cubic interpolation.

In contrast to this, in the embodiment of the present invention, b-spline interpolation or cubic interpolation, which is a three-dimensional interpolation method having excellent interpolation performance among the interpolation methods mainly used for image interpolation in the spatial domain, , We propose a motion prediction method in the frequency domain that extends the above-described shifting matrix algorithm by a quarter pixel unit. In Equation (11), a may be selected from values between 0 and 1, such as 0.25, 0.5, and 0.75.

Accordingly, when a cubic interpolation method such as B-spline or cubic is applied to a shifting matrix, a weighted value is applied using four neighboring pixels, and a predicted block is generated by summing the weighted values. , The vertical shifting matrix (Equations (2) and (2) described above) is developed as Equation (12), Equation (13), and Equation (14).

이 된다. 그리고

일 경우

과

은 4×4의 영행렬을 사용하고,

일 경우

,

가 된다. 그리고 상기 수학식 12 및 수학식 13에서

는 버림 함수를,

는 올림 함수를 나타낸다. 그리고 상기 수학식 14에서, 상기 버림 함수 및 상기 올림 함수에 사용되는 움직임 정보(즉, 실제 움직임에 따른 변위량)로는

대신

가 적용된다.Here, since the motion prediction according to the embodiment of the present invention is 1/4 sub-pixel unit motion prediction,

. And

If

and

Using a 4 × 4 zero matrix,

If

,

. In the equations (12) and (13)

Lt; RTI ID = 0.0 >

Represents a rounding function. In Equation (14), the discontinuance function and the motion information (i.e., the amount of displacement in accordance with the actual motion) used in the up function

instead

Is applied.

인 경우, 1/4 부화소 단위로 볼 때 d는 1/2(즉, 0.5)이 되고,

는 버림 함수 값을 적용할 때 2가 되고,

는 올림 함수 값을 적용할 때 3이 되므로, 수평 시프팅 매트릭스는 아래 수학식 15와 같이 표현될 수 있다.
Thus, for example

, D is 1/2 (that is, 0.5) in the unit of a quarter sub-pixel,

Is 2 when applying the discard function value,

Is 3 when applying the rounding function value, the horizontal shifting matrix can be expressed by Equation (15) below.

또는

인 경우) 참조 블록의 경계 밖에 위치의 주파수 정보를 활용하게 된다. 이때, 주파수 영역으로 변환된 블록에서 특정 위치의 주파수 정보를 가져오기 위해선 특정 위치가 포함되어 있는 블록 전체의 정보가 필요하므로, 불필요한 메모리가 요구된다. 이러한 문제를 해소하기 위하여 본 발명의 실시예에서는 인트라 예측(Intra prediction) 모드 중 vertical 또는 horizontal 모드를 적용하여 필요한 위치를 예측하는 방법을 사용하였다. 이에 관한 예시가 도 5에 도시되어 있다. 여기서, 도 5는 본 발명의 실시예에 따른 시프팅 매트릭스를 이용한 DCT 기반 부화소 정밀도 움직임 예측 방법에서, 인트라 예측 모드에서의 움직임 예측 방식 중 일부를 채용한 예시 도면이다.In the case of the existing bilinear interpolation method, frequency information of two locations is used for subpixel prediction, while frequency information of four locations (ie, four neighboring pixels) is required when b-spline interpolation and cubic interpolation are used. Therefore, when the absolute value of the actual motion displacement amount is smaller than 1 or larger than 3 (for example, in the case of the vertical motion displacement amount,

or

The frequency information of the position outside the boundary of the reference block is utilized. At this time, in order to fetch the frequency information of the specific position in the block transformed into the frequency domain, information of the entire block including the specific position is required, so that unnecessary memory is required. In order to solve this problem, in the embodiment of the present invention, a method of predicting a necessary position by applying a vertical or horizontal mode in Intra prediction mode is used. An example of this is shown in Fig. Here, FIG. 5 is a diagram illustrating a part of a motion prediction method in an intra prediction mode in a DCT-based subpixel precision motion prediction method using a shifting matrix according to an embodiment of the present invention.

가 y축 방향으로

의 움직임이 발생할 경우 탐색 영역의 위쪽 경계 밖의 정보가 필요하게 되며, 이 경우 도 5에서와 같이 탐색 영역의 정보를 위쪽 방향으로 복사함으로써 메모리 소모 문제를 해소할 수 있게 된다. 하기 수학식 16 및 수학식 17은 1/4 화소 단위 움직임 예측에서 도 5에서와 같이 수직 방향으로 예측되는 경우 수직 시프팅 매트릭스의 예를 보여준다.
E.g,

In the y-axis direction

The information outside the upper boundary of the search area is required. In this case, as shown in FIG. 5, the memory consumption problem can be solved by copying the information of the search area in the upward direction. Equation (16) and Equation (17) show an example of a vertical shifting matrix when it is predicted in the vertical direction as shown in FIG. 5 in quarter-pixel motion prediction.

FIGS. 6 and 7 are performance comparison graphs between the motion estimation method and the prior art according to an embodiment of the present invention. Hereinafter, an experiment in which the performance between the motion estimation method according to the embodiment of the present invention and the prior art is compared will be described with reference to Tables 1 to 3 and FIGS. 6 and 7 below.

In this specification, the H.264 / AVC video encoder is used to evaluate the performance of the algorithm (motion estimation method) proposed in the embodiment of the present invention. As a motion prediction method, performance was compared using a full range search method of a quarter pixel unit, and a criterion for determining motion information (i.e., a reference for finding a block having a minimum error) (Sum of Absolute Difference).

크기의 블록으로 공간 영역(Spatial domain)과 주파수 영역(Frequency domain)에서의 움직임 예측에 사용되는 계산량을 나타낸다. 하기 표 1에서 p는 탐색구간, N은 블록 크기를 나타내며, 주파수 영역에서

로 고정된다.
The proposed method of the present invention has an advantage that the IDCT process is not required as compared with the motion prediction in the spatial domain, and the calculation amount is reduced by inducing the recursive equation. Table 1 below

Block represents a calculation amount used for motion prediction in a spatial domain and a frequency domain. In Table 1, p denotes a search interval, N denotes a block size, and in the frequency domain

.

Table 2 and Table 3 below show the results of performing motion estimation in the spatial domain and the frequency domain while varying the quantization parameter (QP) using the first and second frames of the Foreman and Paris images at CIF resolution . Table 2 compares the prediction performance of quarter-pixel motion prediction in the spatial domain and frequency domain according to the QP adjustment using the Foreman sequence. Table 3 shows the performance comparison between the spatial domain and the frequency domain according to the QP adjustment using the Paris sequence. Of the motion prediction performance.

In the case of the Foreman image, the motion prediction method in the frequency domain provides a higher PSNR than the motion prediction method in the spatial domain, but in the case of QP≥21, the total number of bits per frame is higher than the motion prediction in the spatial domain.

In the case of the Paris video, motion prediction in the frequency domain using b-spline interpolation provided a high PSNR regardless of the change in QP, but motion prediction in the frequency domain using cubic interpolation provided spatial domain The PSNR is lower than that of the motion prediction. In the case of b-spline interpolation, the number of bits is QP = 3 and QP≥21, and the cubic interpolation method is higher than QP = 15 and QP≥21. The results of QP except for the specific values show that the frequency domain motion estimation finds more accurate motion information than the motion estimation of the spatial domain, so that the image quality of the image is maintained similar to that of the spatial domain . However, the reason why a large bit rate is generated at a high QP is because the quantization error increases as the QP increases.

6 and 7 show the motion prediction performance in the spatial domain and the frequency domain at QP = 18. Here, FIG. 6 is a graph comparing the performance between the spatial domain using the Foreman CIF video sequence and the motion prediction technique in the frequency domain at QP = 18. And FIG. 7 is a graph comparing the performance between the spatial region using the Paris CIF video sequence and the motion prediction technique in the frequency domain at QP = 18.

Experimental results show that the PSNR of the motion estimation in the frequency domain is higher than that in the spatial domain. In the frequency domain, b-spline (BTD), cubic (CTD), and b-spline (BSD) ), Which are different from the spatial domain. When interpolation of an image is performed in a spatial domain, pixel values of the surrounding pixels are used to generate a pixel, whereas the algorithm proposed in the embodiment of the present invention uses a DCT coefficient. Therefore, it can be seen that even if the PSNR with high motion prediction performance using the specific interpolation method is provided in the spatial domain, the performance is not maintained in the frequency domain.

In the embodiment of the present invention, a quarter-pel motion prediction algorithm in the frequency domain using a cubic function interpolation method is proposed. As a result, as described above, through the simulation using video sequences having various characteristics, the algorithm proposed by the embodiment of the present invention has a higher PSNR than the motion prediction algorithm of the spatial region in the 1/4 pixel unit at most QP values Respectively. Of course, there is a part where the PSNR is lower than the motion prediction in the spatial domain in a specific QP value, but the difference is small enough to be ignored. Also, since the IDCT is omitted and the recursive equation is used, it is possible to maintain a lower computational complexity than the spatial domain, and the structure of the video encoder based on the DCT is simplified. The recently confirmed H.265 uses a variable block with a block size ranging from 4 × 4 to 32 × 32. The motion prediction method in the frequency domain using the shifting matrix will be applicable as described above.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the following claims And changes may be made without departing from the spirit and scope of the invention.

Claims (5)

A motion prediction method for a 1/4 sub-pixel unit for video coding,
Performing DCT (Discrete Cosine Transform) transformation on the input image frame; Performing motion prediction in a frequency domain to acquire a prediction block having a minimum matching error with a current block through block matching between a DCT transformed current image frame and a previous image frame,
In the step of performing motion prediction in the frequency domain, the prediction block may include a horizontal / vertical shifting matrix to which image interpolation is applied in units of quarter pixels using a three-dimensional interpolation function, A matrix operation using a DCT coefficient matrix of four blocks around the prediction block in a search range of a frame,
Wherein the prediction block is defined by the following equations (1) and (2), and the horizontal shifting matrix uses a matrix defined by Equation (3) and Equation (4), and the vertical shifting matrix is expressed by Equation ≪ / RTI >

[Equation 1]

here,

Gt;

Gt; DCT < / RTI >

&Quot; (2) "

Here, f _i is a prediction block

And has a 4x4 block size as any one of the four reference blocks around the reference block,

Wow

Respectively represent the horizontal shifting matrix and the vertical shifting matrix.

&Quot; (3) "

&Quot; (4) "

Where f () denotes the three-dimensional interpolation function, d

Pixel sub-pixel position,

Represents a discarding function,

Represents a rounding function, and D represents

4 x 4 displacement matrix.

&Quot; (5) "

Here, the superscript t represents the transpose matrix, and the discarding function and the rounding function in Equation (5) are the horizontal displacements in Equations (3) and

Instead,

Is applied.
delete The method according to claim 1,
In the matrix operation for obtaining the prediction block,
The displacement matrix of Equation (3), Equation (4) and Equation (5)

The n values of the horizontal displacement amount and the vertical displacement amount,

≪ / RTI >

and

A 4 × 4 zero matrix is used,

≪ / RTI >

And

Using the matrix of the motion vector.
The method according to claim 1,
Wherein the three-dimensional interpolation function is a b-spline interpolation function of Equation (6) or a cubic interpolation function of Equation (7).

&Quot; (6) "

&Quot; (7) "

In Equation (7), a is a selected constant value to be applied to the cubic interpolation function.
The method according to claim 1,
And DCT coefficient information outside the boundary of the search area is required according to the horizontal displacement amount or the vertical displacement amount, the DCT coefficient information in the search area is copied and used for the motion prediction.

Images