Video quality and traffic QoS in learning-based subsampled andreceiver-interpolated video sequences

摘要

Sources of real-time traffic are generally highly unpredictablenwith respect to the instantaneous and average load which they create.nYet such sources will provide a significant portion of traffic in futurennetworks, and will significantly affect the overall performance of andnquality of service. Clearly high levels of compression are desirable asnlong as video quality remains satisfactory, and our research addressesnthis key issue with a novel learning-based approach. We propose the usenof neural networks (NNs) as post-processors for any existing videoncompression scheme. The approach is to interpolate video sequences andncompensate for frames which may have been lost or deliberately dropped.nWe show that deliberately dropping frames will significantly reduce thenamount of offered traffic in the network, and hence the cell lossnprobability and network congestion, while the NN post-processor willnpreserve most of the desired video quality. Dropping frames at thensender or in the network is also a fast way to react to network overloadnand reduce congestion. Our interpolation techniques at the receiver,nincluding neural network-based algorithms, provide output frame ratesnwhich are identical to (or possibly higher than) the original videonsequence's frame rate. The resulting video quality is essentiallynequivalent to the sequence without frame drops, despite the loss of ansignificant fraction of the frames. Experimental evaluation using realnvideo sequences is provided or interpolation with a connectionist NNnusing the backpropagation learning algorithm, the random NN (RNN) in anfeed-forward configuration with its associated learning algorithm, andncubic spline interpolation. The experiments show that when more framesnare being dropped or lost, the RNN performs generally better than thenother techniques in terms of resulting video quality and overallnperformance. When the fraction of dropped frames is small, cubic splinesnoffer better performance. Experimental data shows that thisnreceiver-reconstructed subsampling technique significantly reduces thencell loss rates in an asynchronous transfer mode switch for differentnbuffer sizes and service rates