UC San Diego

© 2014 Springer Science+Business Media New York The cumulative mean squared error (CMSE) is a widely used measure of distortion introduced by a slice loss. We propose a low-complexity and low-delay generalized linear model for predicting CMSE contributed by the loss of individual H.264/AVC encoded video slices. We train the model over a video database by using a combination of video factors that are extracted during the encoding of the current frame, without using any data from future frames in the group of pictures (GOP). We then analyze the accuracy of the CMSE prediction model using cross-validation and correlation coefficients. We prioritize the slices within a GOP based on their predicted CMSE values. The performance of our model is evaluated by applying unequal error protection, using rate compatible punctured convolutional codes, to the prioritized slices over noisy channels. We also demonstrate an application of our slice prioritization by implementing a slice discard scheme, where the slices are dropped from the router when the network experiences congestion. The simulation results show that (i) the slice CMSE prediction model performs well for varying GOP structures, GOP lengths, and encoding bit rates, and (ii) the peak signal-to-noise ratio and video quality metric performance of an unequal error protection algorithm using slices prioritized by the predicted CMSE is similar to that of the measured CMSE values for different videos and channel signal-to-noise. We also extend the GOP-level slice prioritization to frame-level slice prioritization and show its performance over noisy channels.