Artificial Neural Networks for Video Compression via Temporal Interpolation

摘要

In previous work, we have shown that it is possible to achieve high levels of video compression while obtaining good video quality through the use of temporal sub-sampling and interpolation. Temporal sub-sampling was performed by dropping frames from the original video sequence. Remaining frames were then compressed using some form of video compression (e.g. H.261, JPEG frame compression or Adaptive Neural Video Compression). After transmission or storage of these sequences, the original sequence could be reconstructed by first uncompressing the data stream and interpolating the missing frames. Interpolation is performed on a pixel by pixel basis. In previous work we proposed interpolation based on cubic splines. In this paper we propose the use of learning neural networks for interpolation. We then compare the results obtained using cubic splines to those obtained using two types of neural networks, a connexionist neural network using the back-propagation learning algorithm and the Random Neural Network in a feed-forward configuration with its associated learning algorithm. We show that after an appropriate amount of training, these neural networks perform as well or better than the cubic spline techniques.