Distortion propagation modeling and its applications on frame level quantization control for predictive video coding

Abstract

Inter-frame temporal prediction in video coding causes distortion propagation among adjacent frames. As a result, frame level bit allocation and quantization control are intrinsically to be optimized with dependent rate distortion optimization (RDO). Taking inter-frame distortion propagation into consideration, quantization parameter cascading (QPC) is an efficient technique for frame level quantization control in terms of dependent RDO. This paper proposes a temporal distortion propagation model by quantitatively evaluating the temporal distortion dependency. Block-level temporal trajectory is tracked via inter-prediction on down-sampled original frames to construct the temporal analysis chain, and then tree-style dependent analysis is implemented along the trajectory. The amounts of equivalent distortion propagated from the temporally adjacent frames are quantitatively measured. Then, a new distortion model is developed to facilitate measuring the RD (rate distortion) cost in terms of dependent RDO. Finally, a simplified trellis comprised of candidate quantization parameters (Qp) of the frames within one GOP, and the optimal Qp is searched via simplified dynamic programming to achieve global optimization. The simulation results verify that the frame level QPC algorithm achieves 3.46% BD-RATE saving on average, which is contributed by efficient bit allocation along temporal trajectory.