REGION-BASED ADAPTIVE DISTRIBUTED VIDEO CODING CODEC

The recently developed Distributed Video Coding (DVC) is typically suitable for the
applications where the conventional video coding is not feasible because of its
inherent high-complexity encoding. Examples include video surveillance usmg
wireless/wired video sensor network and applications using mobile cameras etc. With
DVC, the complexity is shifted from the encoder to the decoder.
The practical application of DVC is referred to as Wyner-Ziv video coding (WZ)
where an estimate of the original frame called "side information" is generated using
motion compensation at the decoder. The compression is achieved by sending only
that extra information that is needed to correct this estimation. An error-correcting
code is used with the assumption that the estimate is a noisy version of the original
frame and the rate needed is certain amount of the parity bits. The side information is
assumed to have become available at the decoder through a virtual channel. Due to
the limitation of compensation method, the predicted frame, or the side information, is
expected to have varying degrees of success. These limitations stem from locationspecific
non-stationary estimation noise. In order to avoid these, the conventional
video coders, like MPEG, make use of frame partitioning to allocate optimum coder
for each partition and hence achieve better rate-distortion performance. The same,
however, has not been used in DVC as it increases the encoder complexity.
This work proposes partitioning the considered frame into many coding units
(region) where each unit is encoded differently. This partitioning is, however, done at
the decoder while generating the side-information and the region map is sent over to
encoder at very little rate penalty. The partitioning allows allocation of appropriate
DVC coding parameters (virtual channel, rate, and quantizer) to each region. The
resulting regions map is compressed by employing quadtree algorithm and
communicated to the encoder via the feedback channel. The rate control in DVC is
performed by channel coding techniques (turbo codes, LDPC, etc.). The performance
of the channel code depends heavily on the accuracy of virtual channel model that models estimation error for each region. In this work, a turbo code has been used and
an adaptive WZ DVC is designed both in transform domain and in pixel domain. The
transform domain WZ video coding (TDWZ) has distinct superior performance as
compared to the normal Pixel Domain Wyner-Ziv (PDWZ), since it exploits the
'
spatial redundancy during the encoding. The performance evaluations show that the
proposed system is superior to the existing distributed video coding solutions.
Although the, proposed system requires extra bits representing the "regions map" to be
transmitted, fuut still the rate gain is noticeable and it outperforms the state-of-the-art
frame based DVC by 0.6-1.9 dB.
The feedback channel (FC) has the role to adapt the bit rate to the changing
'
statistics between the side infonmation and the frame to be encoded. In the
unidirectional scenario, the encoder must perform the rate control. To correctly
estimate the rate, the encoder must calculate typical side information. However, the
rate cannot be exactly calculated at the encoder, instead it can only be estimated. This
work also prbposes a feedback-free region-based adaptive DVC solution in pixel
domain based on machine learning approach to estimate the side information.
Although the performance evaluations show rate-penalty but it is acceptable
considering the simplicity of the proposed algorithm.
viii