Perceptual classification of MPEG video for Differentiated-Services communications

Abstract

We present a distortion-based packet marking technique for transmission of motion-compensated video over Differentiated Services networks. For each macroblock of an MPEG2 video sequence, the distortion that would be caused at the receiver by its loss is computed. High distortion macroblocks are grouped into perceptually important slices that can be transmitted as premium packets, while lower distortion slices are sent as less expensive, best-effort traffic. Firstly, computation of the distortion introduced in the current frame only is compared to exhaustive computation of the distortion introduced in the entire group of pictures (GOP) due to the error propagation. Secondly, allocation of the premium traffic on a frame-by-frame basis is compared to GOP-wide allocation. Results show that GOP-wide allocation of premium traffic is key in using premium bandwidth efficiently, with strong PSNR gains with respect to the other approaches. We also propose a model-based distortion computation technique, which, combined with GOP-level premium traffic allocation, delivers nearly the same performance of the exhaustive approach at a fraction of its complexity.