Survey of Dirac: A Wavelet Based Video Codec for Multiparty Video Conferencing and Broadcasting - Intelligent Video Event Analysis and Understanding - page 242

SNR scalability in Dirac allows to reconstruct video using fewer numbers of
bands by compromising on video quality. It provides a tradeoff between the
number of bands and video quality. For reconstruction video, the rest of the data is
padding with zero. If all the bands are used, then there is no need of padding zero.
The quality of reconstructed video with lesser number of bands is low as
compared to greater number of bands. The bit stream structure after joining base
layer and enhancement layer1 for both intra and inter frame for SNR scalability
are as shown in Figure 33 (a,b). Note that the header information is necessarily
appended; otherwise the bit stream will not be decoded.
(a)
(b)
Fig. 33 Base layer and enhancement layer1 joining (a) Intra Frame (b) Inter Frame
For spatial scalability, the bit-stream generated by Dirac encoder is splitted into 3
streams same as SNR scalability. At the decoder side, the layers of bit-stream are
combined and the header information for parse unit size of each parse unit is
updated. The size of the reconstructed video and wavelet transform depth and
related information is updated. Note that as oppose to the Figure 33(a,b), there is
no need to append header information for the missing bands as the video is
constructed from the specific number of 4-16 bands and its size is small.
7.2 Computational Complexity Reduction
To reduce the complexity of video encoder and improve the compression
efficiency, fast and efficient motion estimation is crucial as it is the most
expensive component in terms of memory and computational load. The existing
ME search algorithm is too much expensive in terms of time. It uses multiple
hierarchies for all types of inter (both P and B) frames. For example, to encode a
CIF video format, the algorithm requires to generate 4 levels of hierarchies for
both current and reference frames. It computes the optimum motion vector for
each block by calculating SAD using the pattern as shown in Figure 13. Finally,
the search is carried out at the original frame level. Obviously, it computes a huge
amount of SAD calculations that take approximately 80% of the total encoding
time. So, an efficient and faster ME search strategy is required, which could
reduce the encoding time without affecting the accuracy. Efforts are going on to