KR20110033748A - H.264 to mpeg-2 transcoding considering features of variable blocks - Google Patents

Abstract

PURPOSE: A moving picture expert group ii transcoding in a H.264 considering property of a variable block is provided to improve efficiency of a H.264 to MPEG-2 transcoding by effectively advancing motion estimation which occupies the most computational complexity through comparison of a motion vector in a MPEX-2 encoder. CONSTITUTION: A motion vector number about a variable block in an inter mode except a skip of a MBk(M,N) and an intra mode is determined. By using a cost function about a umber of a variable block, a distortion is measured. A motion vector of a variable block having a cost function maximum value among each variable block is determined as a prediction motion vector. Through a ±2 sub pixel search, a motion vector is finally determined. An encoding of an image is performed.

Description

H.264 to MPEG-2 Transcoding Considering Features of Variable Blocks}

Recently, video transcoders have been developed to use multimedia content encoded in the MPEG-2 standard in accordance with the H.264 standard. However, personal video recorders (PVRs) and HDTVs used in the market and at home follow the MPEG-2 standard and do not provide the H.264 standard. As a result, devices that previously supported the MPEG-2 standard cannot play video encoded in the H.264 standard, resulting in a need to purchase a new device. To reduce this cost, we need a transcoder that converts the H.264 standard into the MEPG-2 standard. Therefore, we propose an MPEG-2 transcoder by analyzing the characteristics of H.264 variable blocks.

There are two major research fields of transcoder. First, it is a field that is used as a transcoder between homogeneous and used to transmit the video stream stored in the server to a low-performance client by adjusting the bit rate or screen size. Second, as a heterogeneous transcoder, it is used to convert video standards to other standards in order to transmit video streams to devices with different server specifications. The heterogeneous transcoder is classified into a cascaded pixel domain transcoder (CPDT) [FIG. 1] and a frequency domain based transcoder (DDT) [FIG. 2]. In the case of a serial pixel-based transcoder, a decoder and an encoder are connected in series so that the input image is completely decoded in the YUV format and then re-encoded. As a transcoder, it has the advantage of simplest structure and excellent image quality. However, it has the disadvantage of increasing the amount of calculation by performing the motion compensation operation twice. In the frequency domain based transcoder, the motion prediction operation is performed in the DCT domain. That is, since the process from IDCT to DCT of the decoder is omitted, it shows faster performance than the serial pixel-based transcoder but does not perform IDCT and DCT processes, so errors with reference frames occur, resulting in error accumulation and deterioration of image quality. Cause.

In addition, when transcoding video decoded by the H.264 standard to the MPEG-2 standard, the existing methods differ depending on the intraframe and the interframe [Fig. 3]. First, intraframe transcoding is a brute force method that encodes a decoded H.264 bitstream in accordance with the MPEG-2 standard. Since there is no intra prediction process of the H.264 standard, the MPEG-2 standard has no macroblock mode decision, and the H.264 decoded bitstream is divided into 4 8x8 blocks in 2D 8x8DCT according to the MPEG-2 standard. Encode Secondly, transcoding of an interframe is a method of resampling an optimal motion vector that can be used when encoding motion blocks of a variable block encoded by the H.264 standard to the MPEG-2 standard. At this time, there is a method to find the optimal motion vector by self-motion estimation in MPEG-2 encoder without reusing the motion vector of H.264. However, this method is overloaded because the motion estimation time takes up 60 ~ 80% of the encoder, so most existing methods have a reuse method.

In the present invention, the serial pixel area based transcoder method is used to reduce the deterioration of image quality, and the macroblock mode and the motion transmitted from the H.264 decoder are used to improve the high computational execution speed which is a disadvantage of the serial pixel area based transcoder method. We propose a method of reusing motion vectors in MPEG-2 encoder by grasping the characteristics of vector's surrounding information.

As described in the prior art survey, the average method, the weighted average method, the dynamic search area, etc. were used to determine the predicted motion vector, regardless of the characteristics of the motion vectors of the H.264 decoder. ]. Therefore, in the present invention, in order to predict an optimal motion vector, the SKIP mode and the INTRA mode, which are transmitted from the H.264 decoder to the MPEG-2 coder, are used in the MPEG-2 coder as they are. Use vectors. The structure and contents of the present invention are as follows. First, the motion vectors according to the macroblock mode present in the H.264 decoder, the motion vectors when the full search is performed in the MPEG-2 encoder, and the conventional weighted average method. Compare the motion vectors when applying. Secondly, we suggest the characteristics of moving objects and the characteristics of H.264 motion vector.

H.264 to MPEG-2 transcoding will provide an opportunity to leverage the MPEG-2 infrastructure while providing new types of services. However, because the H.264 and MPEG-2 standards are different, H.264 to MPEG-2 transcoding is required. The present invention proposes to efficiently perform the motion estimation that takes the largest amount of computation in the MPEG-2 encoder for the efficiency of H.264 to MPEG-2 transcoding. Experimental results show that the proposed algorithm has a similar PSNR as compared to the conventional method, but the motion estimation is reduced by about 65%, resulting in excellent performance.

As described in the prior art survey, the average method, the weighted average method, the dynamic search area, etc. were used to determine the predicted motion vector, regardless of the characteristics of the motion vectors of the H.264 decoder. ]. Therefore, in the present invention, in order to predict an optimal motion vector, the SKIP mode and the INTRA mode, which are transmitted from the H.264 decoder to the MPEG-2 coder, are used in the MPEG-2 coder as they are. Use vectors. The structure and contents of the present invention are as follows. First, the motion vectors according to the macroblock mode present in the H.264 decoder, the motion vectors when the full search is performed in the MPEG-2 encoder, and the conventional weighted average method. Compare the motion vectors when applying. Secondly, we suggest the characteristics of moving objects and the characteristics of H.264 motion vector.

(1) Characteristics of H.264 motion vector

5 shows four motion vectors of a variable block of an H.264 decoder and four motion vectors of an MPEG-2 encoder ("Inter MB", "Matched MB",) in order to evaluate the match rate of the motion vectors. The experiment was divided into "MV value matched", "MV value within ± 2", "average value within ± 2"). “Inter MB” expresses the number of macroblocks in the case of INTER mode except SKIP and INTRA mode because MPEG-2 encoder and H.264 encoder mentioned above are different. The number of macroblocks corresponding to the motion vector of the MPEG-2 encoder among the motion vectors existing in the variable block of the .264 decoder. "MV value is within ± 2" is the number of macroblocks in which motion vectors present in the variable block of the H.264 decoder exist within ± 2 of the motion vector of the MPEG-2 encoder. "Average value is within ± 2" is the number of macroblocks in which a motion vector obtained by averaging motion vectors existing in a variable block of an H.264 decoder is within a motion vector ± 2 of an MPEG-2 encoder. According to the characteristics of the experimental images, the BUS, FOREMAN, and COSTGUARD images with constant movements move in the horizontal direction so that the motion vector value is more than 90% within ± 2. However, a large motion image such as STEFAN image does not exist within ± 2.

(2) proposed method

In FIG. 6, when the macroblocks of the H.264 decoder are 63 and 74, the H.264 decoder can be classified into moving objects and blocks indicating a background. That is, a block pointing to a moving object has a clear boundary distinct from the background, and the motion vector of the block having the boundary represents the moving value of the moving object. Therefore, in the present invention, the following method is used to find a block including an interface among candidate blocks as shown in FIG.

Step 1: Determine the number of MBk (M, N)

In this step, the number of motion vectors k for the variable blocks existing in the INTER mode except the SKIP and INTRA modes of MBk (M, N) is determined. However, since M and N are macroblock positions and k is a variable block number, there are at least 1 to 16.

Step 2: determine cost function and predictive motion vector (MVp)

For the number of variable blocks in MBk (M, N), distortion is measured using a cost function as shown in FIG. 8 below.

8 is a distortion value for the top, bottom, left and right of the boundary existing in the kth variable block. m and n are horizontal and vertical lengths of the variable block, and are coordinates of pixels of the current frame. And the cost function is the average of the sum of the distortion values. At this time, the total SAD calculation of the block requires a lot of computation time and does not affect the SAD when the image inside the block is complicated. Therefore, the boundary value in the block is selected to calculate the distortion value. Finally, the motion vector of the variable block having the maximum cost function among each variable block is referred to as a prediction motion vector (MVp). The reason why the cost function value is selected as the maximum value is that since the discontinuity between pixels occurs when the motion spans the boundary of the block, the pixel difference is significantly increased. Therefore, the maximum value is determined by determining that there is an image at the boundary of the block.

Step 3: Determine final motion vector

In the previous step, the motion vector of the variable block having the maximum cost function value is determined as MVp, and finally, the motion vector is determined by performing a ± 2 half-pixel search.

For the experiment environment, VC ++ 2005 was used on PC with Pentium IV CPU 3.06GHz, and the test images were tested at 300 frames each with CIF (352x288) and SIF (352x240) size images. The transcoder was composed of H.264 JM10.2 decoder and MPEG-2 encoder from MPEG Software Simulation Group. In this case, the H.264 decoder decodes the base profile using the H.264 JM 10.2 encoder. The experimental images were encoded at a bit rate of 1 Mbits / sec with a GOP size of 15 frames except B frames. In addition, the quantization parameter is 24, the motion vector search range is ± 15, the reference frame is one, CABAC, RD optimization. The MPEG-2 encoder is identical to the H.264 encoding environment.

On the PSNR side [Fig. 9], 'W-AVERAGE', 'Dynamic window' and 'Proposed method' are 1.7dB on average compared to 'Global search' of MPEG-2 encoder. It decreased by 0.07dB and 0.08dB. In terms of encoding time [Fig. 10], the average decrease was 66.4%, 64.4%, and 64.9%, respectively. Among the existing methods, the 'dynamic window area', which is superior in performance, is almost similar to the 'global search' of MPEG-2 encoders. This is big. The 'CONTAINR' video has almost no motion, so the coding time is almost constant. However, if you look at moving images such as 'FOREMAN' and 'STEFAN' images, 'dynamic window area' does not have a constant amount of calculation from frame 181 to 221, but 'suggested method' performs constant encoding regardless of movement. do.

Claims (5)

Identify the motion vector of H.264 and the motion vector of MPEG-2 baseline. Compares the match rate of the motion vector of H.264 with the motion vector of MPEG-2 encoder (with motion vectors within ± 2). A block that points to a moving object has a clear boundary that is distinct from the background, and the motion vector of the block with this boundary has the movement of the moving object. In order to find the block including the boundary among the candidate blocks, the pixel value and the distortion value within the boundary are measured. If the motion is over the boundary of the block, the discontinuity between pixels occurs, so the pixel difference is significantly larger.