PhD Defense: "Towards Joint Optimality of Spatial, Temporal Prediction and Transform Coding of Video Signals"

Yue Chen

December 3rd (Thursday), 12:00pm
Harold Frank Hall (HFH), Room 1132 (CS Conference Rm)

This work considers joint optimization approaches in video compression, with particular focus on spatial, temporal prediction and transform coding. The main line of research is concerned with a comprehensive model for the correlation between pixels and their spatial/temporal references, and means to effectively exploit multiple sources of information to approach rate-distortion optimal prediction. In intra frame coding, the spatial correlations can be modeled as a 2-D non-separable Markov process, which motivates the derivation of four-tap extrapolation filters that are recursively employed within blocks to predict from nearest neighbors in all available directions. This parametric prediction model can capture the complex phenomenon of decaying correlation with distance from the block boundaries, by careful selection of filter coefficient values, and hence opens the door to tailor the coder to global statistics, by designing a set of filters via an iterative clustering technique. The approach is further extended to inter frame coding by developing spatio-temporal prediction filters, to jointly exploit spatial and temporal references, in contrast with current codecs that switch between spatial and temporal prediction. Moreover, considering the limitation of block-oriented coding, the proposed approach is complemented by an edge-directed joint prediction technique that accounts for the concurrence of objects requiring different predictors which are separated by irregular-shaped motion edges. Another focus of the dissertation is on decoupling transform and prediction block structures. The performance of conventional codecs is undermined by the practical constraint that transform blocks lie within prediction blocks to avoid blocking artifacts that compromise transform efficiency. This shortcoming is circumvented by a pre-filtering technique that is applied to prediction blocks, and thereby makes the decoupled block structure beneficial not only in inter coded blocks but also for compressing heterogeneous residue data. The proposed approaches are integrated within multiple commercial codecs and demonstrated to achieve substantial compression gains.

About Yue Chen:

photo of Yue Chen Yue Chen received the B.S. degree in electronic engineering in 2011 from Tsinghua University, Beijing, China, and the M.Sc. degree in electrical and computer engineering in 2013 from University of California Santa Barbara, where she is pursuing her Ph.D. degree under the guidance of Professor Kenneth Rose at the Signal Compression Lab. In the summer of 2013 and 2014, she interned with the WebM Codec Team at Google Inc., Mountain View, CA. Her research interest includes video coding and transmission.

Hosted by: Professor Kenneth Rose