Region-oriented video coding using the MDL principle and quad-tree optimization

We present a novel approach to find a sub-optimal image segmentation for video coding using a minimum description length (MDL) cost function. We describe the theory necessary to derive an appropriate cost function based on probabilistic models for the image prediction error and motion parameters. A key feature of our approach is the use the MDL cost function to incorporate global bit-rate minimization into the criterion for region segmentation. We apply this theory to the design of a bit-stream level interframe motion-compensated video coding system using a quad-tree optimisation procedure with variable order motion modeling. A computationally feasible approach is illustrated which could potentially be implemented in real-time with appropriate hardware