New methods for motion estimation with applications to low complexity video compression
The goal in video compression is to remove the redundancy in a video sequence while preserving its fidelity. Motion estimation can significantly improve video coding efficiency by reducing temporal redundancy. In this thesis we focus on new methods for motion estimation with applications in low complexity video compression. In this thesis we study the performance of low complexity video coding. Theoretical rate-distortion performance is derived and evaluated for both low complexity video coding and conventional motion-compensated prediction (MCP) based video coding. Based on our analysis of sub-pixel and multi-reference motion search methods used to improve low complexity video coding efficiency, we propose a new refined side motion estimation method to better extract motion information at the video decoder. Motion side estimation methods in general assume a motion model in a video sequence and use computationally intensive motion search. In this thesis we propose a novel side information estimation method based on universal prediction. Our method does not assume an underlying motion model. Instead, it makes predictions based on its observation of the past video information. The experimental results show that this new method can significantly reduce the computational complexity while achieving comparable performance for many video sequences. Lossy image and video compression is often accompanied by annoying artifacts. In this thesis we present a transform domain Markov Random Field model to address the artifact reduction. We present two methods, which we denote as TD-MRF and TSD-MRF, based on this model. We show by objective and subjective comparisons that transform domain post processing methods can substantially reduce the computational complexity compared with conventional spatial domain method and still achieve significant artifact reduction.
Delp, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our