1 EE 5359 H.264 to VC-1 TRANSCODING Vidhya Vijayakumar Student I.D.: 1000-622152 Date: November 3, 20092 H.264 to VC-1 TRANSCODER Objective The objective of the thesis is to implement a H.264 bitstream to VC-1 transcoder for progressive compression. Motivation The high definition video adoption has been growing rapidly for the last five years. The high definition DVD format blue ray has mandated MPEG-2 [3], H.264 [2] and VC-1 [1] as video compression formats. The coexistence of these different video coding standards creates a need for transcoding. As more and more end products use the above standards, transcoding from one format to another adds value to the product’s capability. While there has been recent work on MPEG-2 to H.264 transcoding [3], VC-1 to H.264 transcoding [4], the published work on H.264 to VC-1 transcoding is nearly non-existent. This has created the motivation to develop a transcoder that can efficiently transcode a H.264 bitstream to a VC-1 bitstream. Fig. 1 gives a typical application scenario. Fig. 1 An application scenario for transcoding [39] Details Video transcoding is the operation of converting video from one format to another [5]. A format is defined by characteristics such as bit-rate, spatial resolution etc, as shown in Fig. 2. One of the earliest applications of transcoding is to adapt the bit-rate of a compressed stream to the channel bandwidth for universal multimedia access in all kinds of channels like wireless networks, internet, dial-up networks etc.3 Fig. 2 Transcoding [5] Changes in the characteristics of an encoded stream like bit rate, spatial resolution, quality etc can also be achieved by scalable video coding [5]. However, in cases where the available network bandwidth is insufficient or if it fluctuates with time, it may be difficult to set the base layer bit-rate. In addition, scalable video coding demands additional complexities at both the encoder and the decoder. The basic architecture for converting an H.264 bitstream into a VC-1 elementary stream arises from complete decoding of the H.264 stream and then re-encoding into a VC-1 stream. However, this involves significant computational complexity [6]. Hence there also is a need to transcode at low complexity. Transcoding can in general be implemented in the spatial domain or in the transform domain or in a combination of the two domains. The common transcoding architectures [5] are: Open loop transform domain transcoding Open loop transcoders are computationally efficient, as shown in Fig. 3. They operate in the DCT domain. However they are subject to drift error. Drift error occurs due to rounding, quantization loss and clipping functions. Fig. 3 Open loop transform domain transcoder architecture [5] Cascaded Pixel Domain Architecture (CPDT) This is the most basic transcoding architecture (Fig. 4). The motion vectors from the incoming bit stream are extracted and reused. Thus the complexity of the motion estimation block is eliminated which accounts for 60% of the encoder computation. As compared to the previous architecture, CPDT is drift free. Hence, even though it is slightly more complex, it is suited for heterogeneous transcoding between different standards where the basic parameters like mode decisions, motion vectors etc are to be re-derived.4 Fig. 4 Cascaded pixel domain transcoder architecture [5] Simplified DCT Domain transcoders (SDDT) This transcoder is based on the assumption that DCT, IDCT and motion compensation are linear processes (Fig. 5). This architecture requires that motion compensation be performed in the DCT domain, which is a major computationally intensive operation [3]. For instance, as shown in the Fig. 5, the goal is trying to compute the DCT coefficients of the target block B from the four overlapping blocks B1, B2, B3 and B4. Fig. 5 Simplified transform domain transcoder architecture [5]5 Fig. 6 Transform domain motion compensation illustration [5] Also, clipping functions and rounding operations performed for interpolation in fractional pixel motion compensation lead to a drift in the transcoded video. Cascaded DCT Domain transcoders (CDDT) This is used for spatial/temporal resolution downscaling and other coding parameter changes (Fig. 7). As compared with SDDT, greater flexibility is achieved by introducing another transform domain motion compensation block; however it is far more computationally intensive and requires more memory [3]. It is often applied to downscaling applications where the encoder end memory will not cost much due to downscaled resolution. Fig. 7 Cascaded transform domain transcoder architecture [5] Choice of basic transcoder architecture: DCT domain transcoders have the main drawback that motion compensation in transform domain is very computationally intensive. DCT domain transcoders are also, less flexible as compared to pixel domain transcoders, for instance, the SDDT architecture can only be used for bit rate reduction transcoding. It assumes that the6 spatial and temporal resolutions stay the same and that the output video uses the same frame types, mode decisions and motion vectors as the input video. For H.264 to VC-1 transcoding, it is required to implement several changes in order to accommodate the mismatches between the two standards. For instance, for motion estimation and compensation, H.264 supports 16x16, 16x8, 8x16, 8x8, 8x4, 4x8, 4x4 macroblock partitions (Fig. 8), but VC-1 supports 16x16 and 8x8 only (Fig. 9). The transform size and type (8x8 and 4x4 in H.264 and 8x8, 4x8, 8x4 and 4x4 in VC-1) are different and make transform domain transcoding prohibitively complex. Hence, the use of DCT domain transcoders is not very ideal. Fig. 8 Segmentations of the macroblock for motion compensation in H.264 Top: segmentation of macroblocks, bottom: segmentation of 8x8 partitions [2] Fig. 9 Segmentations of the macroblock for motion compensation in VC-1 [2] From Fig. 10, it can be inferred that, the cascaded pixel domain architecture outperforms the DCT domain transcoders. Also for larger GOP sizes, the drift in DCT domain transcoders becomes more significant.7 Fig. 10 PSNR vs Bit-rate graph for the Foreman sequence transcoded with a GOP size 15, using different transcoding architectures
View Full Document
Unlocking...