KR100986719B1 - computation reduction methode for H.264/AVC motion estimation - Google Patents

Abstract

The present invention relates to a method for reducing the computation amount of H.264 / AVC motion estimation consisting of integer motion estimation (IME) and fractional motion estimation (FME), wherein the search area is adjusted in integer motion estimation, and partial cost calculation is performed in fractional motion estimation. A technical configuration is disclosed in which the amount of computation is greatly reduced by introducing a method.

Integer Motion Estimation, Fractional Motion Estimation, H.264 / AVC

Description

Computation reduction method for H.264 / AVC motion estimation

The present invention relates to a method for reducing the computation amount of H.264 / AVC motion estimation consisting of integer motion estimation (IME) and fractional motion estimation (FME).

In traditional video compression standards such as MPEG-2, MPEG-4, H.264 / AVC, the most computation is consumed by motion estimation. Many such operations arise primarily from iterative block matching operations that compare the reference block with the current block within the search range. In H.264 / AVC, a recent image compressor method, the motion estimation operation is further complicated by employing aggressive techniques such as variable block size motion estimation (VBSME) and quarter-pixel precision motion estimation.

Many studies have been made to reduce the computational amount of motion prediction without significant degradation of compression efficiency. A popular method for reducing the amount of computation is to perform operations on some positions instead of consuming computations on all candidate positions in the search range. To this end, an efficient search pattern is defined based on the fact that the movement is most or near zero in statistical terms. In this manner, the search procedure is initiated at the center of the search window or the estimated motion vector (PMV) and then visits other candidates according to a predefined pattern. However, this method is not suitable for VBSME, which has to derive motion vectors and 41 motion vectors per macroblock for each block. According to this method, only one optimized motion vector is derived. Because it becomes. This approach is also unsuitable for hardware implementations that require simple and intuitive data access to the frame memory for efficient reuse of reading data from the frame memory.

The problem of the present invention is a new motion estimation method employing a global search motion estimation with a search range control system, which reduces the amount of computation from the adjustment of the search area, and at the same time, provides hardware implementation and efficient data reuse due to the nature of global search.

A solution for the above problem is a method for reducing the computation amount of H.264 / AVC motion estimation consisting of integer motion estimation (IME) and fractional motion estimation (FME). Extracting an initial search region; calculating a variance of adjacent motion vectors in the x and y directions; And increasing the size of the initial search window in the initial search area according to the variance value.

According to another aspect of the present invention, there is provided a method for reducing the computation amount of H.264 / AVC motion estimation, which consists of integer motion estimation (IME) and fractional motion estimation (FME). Calculate a sum (SAD), but only four subblocks arranged diagonally in a 16x16 macroblock of 16 4x4 subblocks are used for calculating the SAD.

According to another aspect of the present invention, there is provided a method for reducing the computation amount of H.264 / AVC motion estimation, which consists of integer motion estimation (IME) and fractional motion estimation (FME):

Only eight subblocks alternately arranged in a 16x16 macroblock of 16 4x4 subblocks are used for the SAD calculation.

According to the present invention having the above-mentioned solutions and problems, the operation amount decreases by more than 90% in the IME of the present invention, and in the fractional motion estimation (FME), the operation decreases from 27.8% to 47.1%. The operation is faster, and the basic global search algorithm is adopted, so hardware configuration is easy.

In the motion estimation, a method of reconstructing the motion estimation into integer motion estimation (IME) and fractional motion estimation (FME) is widely used to reduce the complexity of the operation. In this way, integer motion estimation (IME) is used as the starting search point for fractional motion estimation (FME). In this way, the search point of the FME is significantly reduced. In order to further reduce the operation of the FME, a mode predetermination (MPD) scheme and an improved mode predetermination (AMPD) scheme are known. In the present invention, the MPD method is adopted for the FME, and the partial calculation amount measurement method is used in the FME to further improve the speed of the FME.

1. Adjusting the Search Area in the IME

The Joint Model (JM) reference software employs a motion estimation method that starts at the center of the spiral pattern and moves outwards. The optimal motion vector is updated only after the computation amount of all candidates is measured. The best candidate is updated only when the computation amount of the current candidate is less than the computation amount of the previous best candidate. If not, the best candidates remain unchanged. In order to reduce the calculation required for the calculation of the amount of calculation, the calculation of the amount of calculation is immediately stopped if the amount of calculation being measured at this time is larger than the amount of calculation of the immediately best candidate. In order to take advantage of the early termination of the measurement, it is desirable to start the motion estimation at the position where the final motion vector is likely to be the highest. Due to the spatial correlation of the image, the probability of getting very close to the estimated motion vector (PMV) final motion vector is very high. Therefore, PMV is a good candidate as a starting position of motion estimation that reduces computation through early termination. Conventional Method 1 (Tung-Chien Chen, Yu-Han Chen, Sung-Fang Tsai, Shao-Yi Chien, and Liang-Gee Chen, "Fast Algorithm and Architecture Design of Low-Power Integer Motion Estimation for H.264 / AVC, Fast Transition Estimation with 4SS Algorithm for IEEE Trans. Circuits Syst. Video Technol., Vol. 17, No. 5, pp. 568-577, May 200) ME) algorithm is proposed, but it is not suitable for variable block size motion estimation (VBSME) adopted in H.264 / AVC because 4SS finds only one optimal motion vector (MV). In order to overcome this limitation of the 4SS algorithm, the conventional method 1 proposed a method of starting a motion estimation (ME) at a plurality of start positions including PMVs and motion vectors (MVs) of adjacent macroblocks.

1 (a) is a block diagram showing the search range of the IME of the conventional method 1, Figure 1 (b) is a block diagram showing the search range of the IME of the present invention.

In FIG. 1 (a), the large white squares have (x, y) coordinates of four vertices (-16, -16), (-16,15), (15, -16), and (15,15), respectively. Shows the search window. The search window is the maximum searchable range. The inner small shaded rectangle is derived from adjacent macroblocks of the motion vectors, and the dotted rectangle outside the shaded rectangle is derived by adding an experimentally selected parameter to the shaded rectangle. The small triangle shows the possible initial search points, which are the four vertices of the dotted rectangle, the PMV, and the coordinate (0.0) point. The 4SS algorithm is performed with the above initial search points to obtain motion vectors (MVs) for various block sizes.

The fast motion estimation algorithm of the conventional method 1 significantly reduces the search points compared with the global search algorithm. This reduction is more apparent when the motion vectors of adjacent macroblocks are near each other. In such a case, multiplexing the 4SS algorithm at different starting points is likely to generate overlapping search points because these starting points are adjacent to each other. Thus, such repetitive operations can be exhaustive. In the present invention, a new IME algorithm employing a global search method for adjusting the search area is applied. Global search eliminates wasted computation and is very efficient for hardware implementation due to the conventional algorithmic structure. However, global search consumes a lot of computation because it performs calculations on all candidate points in the search window. Figure 1 (b) shows the adjusted search range of the present invention. The search areas indicated by the shade in FIG. 1 (a) are extended by ΔX and ΔY in the horizontal and vertical directions, respectively. ΔX and ΔY are obtained from the variance of adjacent motion vectors. If adjacent motion vectors exist close to each other, the motion vector of the current macroblock may also exist near these adjacent motion vectors. Therefore, the search range adjusted by the small values ΔX and ΔY is sufficient to include the motion vector of the current macroblock. On the other hand, if the variance of adjacent motion vectors is large, it is difficult to estimate the current motion vector. Therefore, it is necessary to increase ΔX and ΔY so that a wide area may be searched so that the motion estimation does not miss the optimal motion vector.

Hereinafter, an algorithm for adjusting the search area of the present invention will be described.

2 is a block diagram illustrating motion vectors of adjacent macroblocks of the present invention.

Step 1: Set initial x and y coordinates as follows:

xmin = min (MVA_x, MVB_x, MVC_x, MVD_x)

xmax = max (MVA_x, MVB_x, MVC_x, MVD_x)

ymin = min (MVA_y, MVB_y, MVC_y, MVD_y)

ymax = max (MVA_y, MVB_y, MVC_y, MVD_y)

Step 2: Calculate four vertices of the initial search range: (xmin, ymin), (xmin, ymax), (xmax, ymin), and (xmax, ymax).

Step 3: Calculate the variance of adjacent motion vectors in the x and y directions:

xver = (12 (xmax xmin) + 4 (ymax ymin)) / 4

yver = = (12 (ymax ymin) + 4 (xmax xmin)) / 4

Step 4: Using xver and yver from Step 3, find the number of additional lines in Table 1 from ΔX and ΔY, that is, in the x and y directions.

Step 5: Select the final search area by adding the four vertices from Step 2 to the additional lines from Step 4. The four vertices of the new search range are (xmin-ΔX, ymin-ΔY), (xmin-ΔX, ymax + ΔY), (xmax + ΔX, ymin-ΔY), and (xmax + ΔX). , ymax + ΔY).

In steps 1 and 2, (xmin, ymin), (xmin, ymax), (xmax, ymin), and (xmax, ymax) are obtained from MVA, MVB, MVC, and MVD as four adjacent MVs. These are the motion vectors (MVs) of adjacent macroblocks, as shown in FIG. 2, in which the shaded squares indicate the current macroblock and the four white squares indicate adjacent macroblocks. In Step 1, MVA_x and MVA_y indicate the x and y coordinates of MVA, respectively. The same notation applies to other motion vectors (MVs). As a result of this algorithm, the search area increases as the variance of adjacent motion vectors increases. This is a valid result because motion vector prediction becomes difficult as the variance of adjacent motion vectors increases.

Table 1 shows the relationship between the search range and the variance values. The figures in Table 1 are experimentally determined values. All operations of the above algorithm can be performed by shift and add operations. Therefore, if the above algorithm is applied, hardware can be implemented at a very low cost.

Xver △ X Yver △ Y 0 2 0 One One 3 One 2 2 3 2 2 3 3 3 2 4 4 4 3 5 4 5 3 6 4 6 3 7 4 7 3

In the present invention, the average number of search points for one macroblock is slightly increased compared to the conventional method 1, but is superior to the conventional algorithm in view of the global search method. Compared with the global search algorithm which has (-16, +15) search areas in both x and y directions, the proposed search area reduces the computation amount by more than 90%.

3. Partial cost evaluation for FME

Since the minimum block size for FME is 4x4, the hardware implementation of FME consists of designing a 4x4 block computation engine and using the engine repeatedly to calculate the entire 16x16 block. In this invention, the FME engine also adopts this method in hardware implementation, and further proposes a new method for reducing the amount of computation. In order to find the best matching position, the FME engine calculates the sum of absolute differences (SAD) between the current block and the reference block. Differences in all the pixels in the blocks are calculated for the calculation of the SAD. To reduce this computation, the present invention applies a new algorithm that uses only subsets of the pixels in the block instead of using all the pixels.

3 is a schematic diagram showing subsets of pixels used for the calculation of the SAD of the present invention.

In FIG. 3, the large squares represent 16x16 macroblocks and the small squares represent 4x4 blocks. The shaded squares indicate subsets of the pixels used for SAD calculation. In FIG. 3 (a), only four 4x4 blocks arranged diagonally in a 16x16 macroblock are used for SAD calculation, that is, only 64 pixels are used. As a result, the amount of computation is reduced to 25%. This partial operation may not accurately reflect the movement, which may result in deterioration of the compression efficiency. 3 (b) shows another example of partial SAD calculation. In this example, eight 4x4 blocks that are alternately placed in a 16x16 macroblock are used for SAD operations, and the overall amount of computation is reduced by 50%. In this case, the calculation amount is larger than that in Fig. 3 (a), but the compression efficiency is better than that in Fig. 3 (a). Therefore, one of two methods shown in FIGS. 3 (a) and 3 (b) may be selectively used.

When applying the method shown in FIG. see.

1 (a) is a block diagram showing the search range of the conventional IME, Figure 1 (b) is a block diagram showing the search range of the IME of the present invention.

2 is a block diagram illustrating motion vectors of adjacent macroblocks of the present invention.

3 is a schematic diagram showing subsets of pixels used for the calculation of the SAD of the present invention.

Claims (5)

delete In the method of reducing the amount of computation of H.264 / AVC motion estimation consisting of integer motion estimation (IME) and fractional motion estimation (FME): In the integer motion estimation, when motion vectors of a macro block adjacent to a current macro block are referred to as MVA, MVB, MVC, and MVD, Step 1: Set initial x and y coordinates to the following equation xmin = min (MVA_x, MVB_x, MVC_x, MVD_x) xmax = max (MVA_x, MVB_x, MVC_x, MVD_x) ymin = min (MVA_y, MVB_y, MVC_y, MVD_y) ymax = max (MVA_y, MVB_y, MVC_y, MVD_y) Step 2: Set up the initial search range with four vertices: (xmin, ymin), (xmin, ymax), (xmax, ymin), and (xmax, ymax) Step 3: Calculate the variance xver, yver of adjacent motion vectors in the x and y directions by the following equation (2) xver = (12 (xmax xmin) + 4 (ymax ymin)) / 4 yver = (12 (ymax ymin) + 4 (xmax xmin)) / 4 Step 4: Extract ΔX and ΔY values from the table where ΔX and ΔY are determined according to xver and yver obtained in Step 3 Step 5: Final search areas (xmin-ΔX, ymin-ΔY), (xmin-ΔX, ymax + ΔY) by adding the four vertices obtained in Step 2 to the ΔX and ΔY values obtained in Step 4 , (xmax + ΔX, ymin-ΔY), and (xmax + ΔX, ymax + ΔY) Computing amount of H.264 / AVC motion estimation method comprising a. The method according to claim 2, wherein ΔX and ΔY increase as the xver and yver increase, respectively. delete delete

Images