PhD Defense: "Redistributed Video Codec Complexity with Relaxed Latency Constraints"

Malavika Bhaskaranand

September 10th (Tuesday), 1:00pm
Harold Frank Hall (HFH), Room 4164

In this dissertation, we develop a video compression scheme that is suited for applications such as UAV video surveillance where the encoder complexity needs to be low, but the decoder complexity can be high. We propose an encoder that uses the global motion information available in UAVs for global motion compensated prediction. This encoder is superior to a complexity-constrained H.264/AVC encoder because it does not need to transmit motion vectors and it replaces the highly complex block motion estimation engine with the relatively simpler global motion compensation. The adoption of global motion compensation in place of the block motion estimation in standard H.264/AVC encoders saves about 40–45% of the bits required to achieve a given frame quality or equivalently achieves BD-PSNR improvements of 1.7–2.5 dB. We also propose a second low complexity encoder with global motion compensation and quantization matrices designed using a spectral entropy based bit allocation scheme that achieves near constant PSNR within groups of pictures at the cost of small increases in delay and complexity, and a small drop in compression efficiency. This encoder has lower complexity and requires 30% fewer bits to achieve the same frame quality as a complexity-constrained H.264/AVC encoder.

In addition, we design a high complexity decoder that takes advantage of the computational resources and relaxed latency constraints at the ground station to further improve the performance gains of the proposed encoders for non-real-time applications. This decoder utilizes the correlation across frames through a colored observation noise Kalman filter applied along motion trajectories on the transform coefficients reconstructed by the “matched” decoder and improves the reconstructed frame quality by 0.2–0.5 dB on average. The proposed compression system consisting of the low complexity encoder and the high complexity decoder achieves about 50% bitrate savings or equivalently 2–3 dB BD-PSNR improvement over the complexity-constrained H.264/AVC encoder with a standard-prescribed low complexity decoder. Our proposed approach also provides substantial bitrate savings of 65% or equivalently 4 dB BD-PSNR improvement over the Wyner-Ziv DISCOVER codec, which also has a low complexity encoder and a high complexity decoder. Further, our high complexity decoder can be paired with a low complexity H.264/AVC encoder with motion estimation constrained to 8×8 blocks and half pixel accuracy, to improve the frame reconstruction quality by 0.2–0.6 dB on average.

About Malavika Bhaskaranand:

photo of malavika bhaskaranand Malavika Bhaskaranand received her B.E. degree in 2004 from the National Institute of Technology Karnataka, India. She was a Senior Design Engineer at Ittiam Systems Pvt. Ltd., India till 2007 where she worked on developing MPEG2 video codecs and an MPEG2-to-H.264 transcoder for TI DSP processors with applications to personal entertainment and broadcast systems. In 2008, she obtained her M.S. degree in Electrical and Computer Engineering from UCSB, where she is currently working towards her Ph.D. under the guidance of Prof. Jerry Gibson. She also interned at Thomson Corporate Research (now Technicolor Research & Innovation) at Princeton, NJ in the summer of 2009. She has published 8 papers in peer-reviewed conferences and has filed for 2 patents. Her research interests include video processing and compression.

Hosted by: Professor Jerry Gibson