英文翻译参考

毕业设计（论文）外文参考资料及译文译文题目：学生姓名：学号：专业：所在学院：指导教师：陈正宇职称：讲师、工程师2011年12 月8 日H.264 Real Time Implementation on DSP——摘自《International BroadcastingConvention 2003》第488页~第499页A. Dueñas1, J. C. Pujol2, A. Martín2, C. Peláez2, F. Díaz2, G.Gomez1, F. Martin11PRODYS, Spain; 2Universidad CIII de Madrid, Spain AbstractH.264/MPEG-4 A VC is the latest generation standard from the ITU-T and ISO/IEC MPEG standardization bodies and it builds on the success of the previous ITU-T and MPEG standards. H.264/MPEG-4 A VC offers major advantages compared with previous standards, among them its high compression efficiency, although it demands a high degree of computational complexity to be able to make use of the advantages of its new techniques. In this paper several implementation issues of the H.264/MPEG-4 A VC standard have been reviewed, dealing with some of the real time challenges that appear on this kind of implementation. This implementation is to be executed on a general purpose DSP conforming with the requirements of the broadcast environments and should provide very important advantages over an MPEG-2 Encoder system.IntroductionThe recently standardized H.264/MPEG-4 A VC video coder (1) (formerly known as ITU-T H.26L) is the result of the work carried out by a Joint Video Team (JVT) part of the International Telecommunication Union (ITU-T VCEG –Video Coding Experts Group) and of the International Organization for Standardization (ISO/IEC MPEG– Moving Picture Experts Group). This version of the ITU-T H.26x standards is added, as well as Part 10 of the MPEG-4 Standard (also called MPEG-4 Advanced Video Coding) and, due to the reputation and experience of these two organizations in this field and the performance results obtained by this new coder, this solution is called to play an important role in the Broadcasting market place.This standard, born under a ‘clean slate’ of ‘back to basics’ paradigm, is designed to enhance the compression efficiency (even at very low bit rates) compared to existing video compression coders. It aims at covering a wide variety of applications ranging from digital storage media, television broadcasting, Internet streaming to real-time audiovisual communication. Typically, a distinction between ‘conversational’ and ‘not conversational’ applications is established in order to impose the appropriate delay constraints.Another important characteristic of this standard is that two abstraction layers are sharply separated: a Video Coding Layer (VCL) and a Network Adaptation Layer (NAL). This fact denotes the new standard vocation of being totally network-agnostic. Besides, special emphasis is placed in recovering from the errors introduced by the transmission networks (see, for example Stockhammer (3) or Sullivan (4)).Broadcasting is one of the applications where this new standard is expected to have special relevance. However, real-time performance and efficient DSP implementations must be accomplished to achieve its potential enhanced performance, yet maintaining an acceptable degree of computational complexity. In our preliminary experiments we have focused on the baseline profile, though our main interest is in broadcasting, which will lead us shortly to the consideration of the characteristics ofthe main profile.The remainder of this paper is organized as follows. The next section is devoted to the analysis of the main technical characteristics of the new coder, stressing its new features with respect to previous standards. Following that, a preliminary evaluation of the key features in terms of its rate-distortion characteristic versus its computational cost is presented. Finally, an assessment of the performance of some of these techniques is offered. Some conclusions and further directions close this paper.H.264/MPEG-4 A VC technical featuresIn the following sections a brief description of the most relevant novelties of the H.264 standard will be presented. We will focus on those algorithmic features supported by the baseline profile coder. One of its most important characteristics is the great range of different possible implementations of the basic components allowed in the coder that result from the number of prediction modes, picture subdivisions or motion vector references allowed. A real time DSP implementation of such a great degree of freedom requires a careful evaluation of efficient techniques for the determination of the optimal parameters in terms of rate-distortion efficiency and computational cost requirements. Next, we will review some of these options.Intra prediction modesIn contrast to previous video coding standards, spatial prediction is always performed in H.264/MPEG-4 A VC. Besides, two different segmentation modes are allowed: Intra_4 ×4 and Intra_16 ×16. When using the first one, nine prediction modes are available for the luminance component: eight directional prediction modes plus one DC. The Intra_16 ×16 mode, intended for smooth areas, provides four luminance prediction modes (three directional and one DC). Chrominance information can also be predicted using four different modes (1 DC plus 3 directional).As can be imagined, an exhaustive search for the most convenient mode results in a significant computational burden and a DSP implementation requires an efficient algorithm to perform this operation.Motion CompensationMotion compensation is another building block in which a plethora of prediction modes are available. Thus, each P-type macroblock can be partitioned into several block sizes and shapes (rectangular shapes are also allowed together with the traditional square forms) to allow an accurate description of motion. Initially, 16 ×16, 16 ×8, 8 ×16 and 8 ×8 blocks are available, and there exists the possibility of further dividing the last one into 8 ×4, 4 ×8 and 4 ×4 blocks or coding it in Intra_4 ×4 mode. As a consequence of this large number of modes, a balance between the number of motion vectors employed and the benefits obtained must be reached, not to mention the need for computationally efficient algorithms to guide the search of the appropriate solution.Unrestricted motion vectors and half- and quarter- sample accuracy are permitted. Besides, multiple frame references are also allowed, requiring the storage of these reference frames both in the coder and decoder ends. Therefore, a smart buffer management system must be designed to cope with that complexity efficiently.De-blocking filteringThe perception of the block segmentation boundaries can be one of the most noticeable artifacts present in a coded video sequence and consequently the mitigation of this effect is extremely important. A conditional smoothing filter is applied to the edges of each macroblock with 5 degrees of smoothness (including a no-filtering degree) that depend on the type of content present on the filtered and adjacent macroblocks. It is worth noting that this filter, in contrast with previous video coders, is implemented inside the coding loop and therefore its effects are taken into account in the calculation of the coded picture and included on the bit stream.Transform coding and coefficient quantificationSimultaneous scaling and transformation of each macroblock is performed using 4×4 DCT transforms where a distinction between Intra_16×16 DC luminance coefficients and the rest of the 4×4 luminance and AC chrominance components of the remaining modes is established. Chrominance DC coefficients, however, employ a 2×2 DCT transform.Non-linear scalar quantification is performed following a zig-zag scan order (an alternative field scan is available for field parameters). Both for chrominance and luminance parameters, there are 52 available quantification steps designed in such a way that there is an increase in the factor used for scaling of approximately 12% in consecutive quantification steps. The usage of this new type of fix-point transform is beneficial for a fix-point DSP implementation, because its implementation will be fully accurate and will remove the possibility of the existence of mismatch on the transforms used by the different codec implementations.Entropy codingA Universal Variable Length Coding (UVLC) method using Exp-Golomb codes is employed for all syntax elements except for the quantized transform coefficients. For these last coefficients two entropy coding methods are on hand in this coder: CA VLC (Context Adaptive Variable Length Coding) or CABAC (Context Adaptive Binary Arithmetic Coding). However, in the baseline profile only CA VLC is supported and CABAC is added to the Main Profile. These new methods, primarily CABAC, increase substantially the complexity of the algorithms and are difficult to perform on the DSP selected to implement this algorithm.Performance evaluationGeneral descriptionTo test the performance of the new features of H.264 some well known video sequences have been used: ‘foreman’ (QCIF format), ‘football’ (CIF), ‘tempete’ (CIF) and ‘paris’ (CIF).The software used for all the experiments is based on the JM61d version of the H.264/A VC reference software.We have performed two main types of test on a Pentium CPU: an assessment of the potential benefits of the de-blocking filter in combination with different quantization parameters and an evaluation of those obtainable by using multiple reference frames and macroblock segmentations. A 1 GHz Pentium III with 128 MB of RAM has been used for the former experiments and a 2 GHz Pentium 4 with 256 MB for the later. Some further DSP testing was performed on a Texas InstrumentsT MS320DM642™ at 600 MHz DSP (5) (6).DSP Device ArchitectureThe DSP implementation described in the present paper have been developed on general purpose Digital Signal Processors. The TMS320DM642™ device from Texas Instruments has been chosen for this implementation; this has Very Long Instruction Word (VLIW) processor core, with a highly parallel and deterministic architecture with eight functional units. This is a high performance fixed-point 600 MHz clocked Digital Signal Processor capable of providing 4,800 MIPS, that has a 16 + 16 Kbytes of L1 of cache memory and 256 Kbytes L2 of cache memory and a 64-Bit External Memory Interface (EMIF) and Enhanced Direct-Memory-Access (EDMA) Controller with 64 Independent Channels. This device also adds a set of video ports to simplify the acquisition and display of the video signals.De–blocking filter and QuantizationIn this section some experimental results are presented to assess the effectiveness of de-blocking filtering. For motion estimation in P-planes, blocks sizes of 16 ×16, 16 ×8, 8 ×16 and 8 ×8 have been used. A single slice per plane is used. In addition, the Flexible Macroblock Order is disabled. Skipped frames are not allowed and motion estimation is performed using 5 reference images.To experimentally assess the de-blocking filter, it has been selectively enabled or disabled.Besides, the experiments have been carried out for several values of the quantization parameter (QP). The finer the quantization is, the smaller the QP parameter used. Consequently, the images are more faithful to the original ones. For our experiments, we have used the following values of the de-blocking filter parameters: for the offset to access the α and C0/2 tables we have used ‘–2’, while the chosen value for the offset to access the β tables has been ‘–1’.Now that the experimental setup has been described, we proceed to present the results obtained. We have conducted three types of measurements: average coding time per image, conducting bit rate and SNR of the luminance component. With respect sequences.，we have selected two of them to illustrate this graphically (see Figure 1).Figure 1: Mean Coding Time per frame vs. the parameter of quantification QPAs it can be observed, the use of the de-blocking filter does not entail aconsiderable increase the coding time. For example, the increase for the ‘football’ sequence is around 3%. If a coarse quantization is used (QP> = 40), the increase in the coding time is insignificant (for QP higher than 40 the coding time increase is negligible). Once it has been shown that one of the potential disadvantages of the de-blocking filter is not very significant, we will evaluate its potential advantages.It is expected that the inclusion of the de-blocking filter would allow a reduction in the resulting bit rate, since it aims at eliminating the artificial high frequencies that appear in the block borders (edge effects), providing smoother images. Results concer ning bit rate and luminance SNR for the ‘football’ sequence are illustrated in Figure 2. Very similar results are obtained for other video sequences.Figure 2: (a) Resulting Bit Rate vs. QP (b) Luminance SNR vs. QP As it becomes obvious, the obtained bit rate reduction is negligible for fine quantization (0.1% for QP = 30); however, the coarser the quantization becomes, the higher the bit rate reduction obtained (from 5% to 8.5% for QP = 40).With respect to the SNR, the results do not seem to be good. Nevertheless, perceptually, the use of the de-blocking filters turns out very effective in coarse quantization cases (QP higher than 40).Summarizing, the convenience of using the de-blocking filter depends on the application. Besides, this technique can be implemented very efficiently on the selected DSP, because this is a method that can be parallelized very efficiently and its increment on the execution time will be minimal, offering potential benefits (bit rate reduction and quality increase) primarily in coarsely quantized videos. As we have already mentioned, the inclusion of the de-blocking filter is optional and therefore, it can be disabled in those cases where it is considered that it may introduce a degradation in the perceived quality of the image.Multiple reference frames and macroblock segmentation Since multiple reference frames and macroblock segmentation are closely related techniques, they have been simultaneously tested to evaluate their overall performance. Firstly, we show how multiple reference frames can improve coding efficiency and secondly some video sequences are processed combining multiple reference frames and several subdivision parameters.Fully exhaustive motion estimation is used on ‘foreman’, ‘football’ and ‘paris’which are divided into 12 images per subsequence. For the encoding of the actual image, ‘N’, three previous images, ‘N-1’, ‘N-2’, ‘N-3’, and the first image have beentested as reference frames. The macroblock size is set to 8 ×8 and a search range of [-32,32] is used. For this test, we have evaluated the MV rate that points to a certain reference frame.Figure 3: Rate of MV that point to the different reference frames Figure 3 depicts the results for ‘foreman’ and ‘football’ (400 frames have been coded for ‘foreman’ and 90 for ‘football’). As can be seen, the use of 3 reference images is advantageous. Even for the ‘football’ sequence, which presents a great amount of motion, there is still a non-negligible percentage of MVs that point to the third reference frame. However, as could have been imagined, this rate decreases along with the distance of the reference frame from the current frame.Having established the utility of multiple reference frames, we have encoded ‘paris’, ‘tempete’ and ‘football’ using different macroblock segmentations. In order to make the notation briefer, let us consider a maximum macroblock subdivision of 16 ×16 named with ‘10000’; ‘11000’ for macroblocks of 16 ×8 or 8 ×16; ‘11100’ is used if a subdivision of 8 ×8 is allowed, ‘11110’ for8 ×4 or 4 ×8 and ‘11111’ for any other macroblocksize.The sequences have beencoded using all the combinationsfrom 1 to 4 reference frames(always the adjacent previousones), ‘10000’to ‘11111’macroblock subdivision anddifferent QP (25, 35 and 45). Thecommon settings are: sequencetype of ‘IPPP’, no search rangerestrictions, disabled de-blockingfilter and no RD-optimized modedecision.Because of the huge amountof data generated, only the mostrelevant results are presented.FFigure 4: Average coding time per sequence Let us first show the time taken to code every single sequence. Since the reference software uses fully exhaustive motion search, the time spent is almost the same for every combination of coding settings, with a negligible variance. Underthese conditions, time has nearly a linear behavior. Though this might be seen as obvious, it is worth noting that smart algorithms able to provide a good estimate of the optimal combination of these two parameters can considerably change this performance, but the usage of these new techniques will always substantially increase the complexity of the coder and a careful evaluation of the advantages is needed. The DSP implementation may show this linear behavior, but the motion estimation algorithm is efficiently performed on the DSP processors selected. This is because of its parallel nature and the efficiency of this DSP architecture to perform the core operations (SAD) present on the motion vector search method. Moreover, the algorithm selected on the DSP implementation is not a full search motion estimation but a more refined approach called Distribute Tornado Search that obtains similar results to full search with a fraction of the complexity. This reduces the portion of the overall processing time needed for the motion estimation and the increase on the coding time no longer shows a linear behavior. It is also important to note that the usage of multiple reference frames not only increases the number of operations required but also increases the memory requirements on the processor. This may be a limiting factor on DSP implementations where the size of the available memory may be an issue when we want to use a high number of multiple reference frames.It is important to emphasize that the experiments have been configured in a way that for a given video sequence the SNR remains almost constant irrespective of the number of reference frames and the macroblock segmentation used. This enables us to make comparisons in terms of the compression rate obtained for a given quality.Figure 5 shows the bit rate obtained for ‘football’ and ‘paris’ which have exhibited very different behaviors. It is worth noting that the best compression in ‘football’ is achieved when more reference frames are used, rather than smaller macroblock subdivision. On the other hand, ‘paris’ is more highly compressed if a smaller size of macroblock is allowed. These results lead us to choose a balance between number of reference images and minimum macroblock sizes: a diagonal move in the grid of coordinates is the most sensible choice if the nature of the video is unknown.Figure 5: Bit rateIf we focus on ‘paris’, two reference frames and ‘11000’ macroblock size, we get 14% better compression with 64% more time, for the same SNR. Three reference frames and ‘11100’ achieve 18% lower bit rate, spending 160% of the time used by the encoder with one reference frame and 16 ×16 macroblock size.It is worth mentioning that, instead of a strictly decreasing function, we find a local minimum at ‘11110’. As our final aim is the maximization of the compression rate, which we cannot do directly, we pursue the minimization of the prediction error (which most of the times redounds in a maximum compression rate). However, thereare situations when both objectives do not converge uniformly. It is therefore worth noting that when using 4 ×4 macroblocks size, we need to pay close attention to the cost function we want to minimize.ConclusionsIn this paper, some of the new features of the ITU-T H.264/MPEG-4 A VC standard have been evaluated, such as de-blocking filtering, multiple reference frames and macroblock subdivision for the motion estimation. Compression performance, quality and coding time required have been analyzed for different sequences.The type of application and the necessities of the user will be crucial to decide on the convenience of using the de-blocking filter. Its impact on the coding time is not really very substantial and, when using a fine quantification, the benefits of its use are not very noticeable. Nevertheless if bandwidth restrictions are imposed, and therefore quality is not high, the de-blocking filter is able to reduce bandwidth around 5% and the subjective quality of the video obtained becomes considerably better.It has been shown that a good combination of multiple reference frames and macroblock segmentation can increase the compression efficiency of the coder, increasing however substantially the processing time required to perform the complete motion estimation.The processing power limitations found on the current DSP devices will make it necessary to consider a number of tradeoffs to be able to develop these techniques in real time in commercial products. Advanced algorithms need to be used to perform the multiple reference frames search method as well as the macroblock segmentation. Also they will be utilized to take advantage of what these techniques offer without increasing significantly the required processing time of the complete coder to the point of making it unsuitable for real time implementations in the current DSP devices.H.264在DSP上的实时实现——摘自《国际广播节目大会2003》第488页~第499页A. Dueñas1, J. C. Pujol2, A. Martín2, C. Peláez2, F. Díaz2, G.Gomez1, F. Martin11PRODYS, Spain; 2Universidad CIII de Madrid, Spain摘要H.264/MPEG-4 AVC是由ITU - T的和ISO / IEC的MPEG标准化机构提出的最新一代的标准，它建立在以前的ITU - T的和MPEG标准成功的基础上。