Model-based distortion estimation for perceptual classification of video packets

In video communications over IP networks, quality of service (QoS) guarantees must be introduced to limit the effect of packet losses. In particular, end-to-end QoS can be improved if packets are protected according to the distortion that would be introduced at the receiver by their loss. In the traditional analysis-by-synthesis (AbS) approach, each packet is assumed lost, error concealment applied, the sequence decoded, and the resulting overall distortion computed. This process produces reliable distortion estimates, but is computationally demanding. In this work, we present a hybrid approach: the distortion introduced in the current frame is evaluated with the AbS method, while the distortion in future frames is estimated by means of a statistical error-propagation model. Results obtained on eight, widely different H.264 sequences show that the proposed model successfully estimates overall distortion with very low complexity. Network simulations also show that model-based packet classification, when used for video transmission over DiffServ networks, delivers PSNR results which are consistently within 0.1 dB compared to the AbS technique.