Error resilient compression of digital video data

The combination of compressed video data to be communicated to and from a mobile user poses challenging problems. The Mobile Multi-media Communication project of the Delft University of Technology tries to find solutions to these problems. There are five research disciplines involved: the application, the user interface, the video compression, the network protocol and the transmission.

For video compression, the problem which is specific for mobile video communication is the presence of transmission errors that corrupt the compressed bit stream. The standard compression techniques are not sufficiently equipped to deal with time errors. Therefore, compression techniques are being developed today to reduce the impact of errors. In tins thesis we present new approaches.

We have modified a H.263 compression standard to make it more error resilient Instead of using the previous frame for prediction, a prediction frame is used which is composed of several previous frames which smoothes the impact of errors. This technique is not dependent on knowledge whether or where any errors have occurred. Also an error detection algorithm is applied which searches in the decoded reconstructed frame for errors. The locations of the errors are signalled to the encoder which intracodes at that position a macro block, some frames later due to the delay In the meantime at the decoder error concealment is applied. Results show a substantial improvement in visual quality at the same bit rate.

The trade-off between this approach and an error resilient network protocol has also been addressed. An experimental approach was chosen. We looked at the effect of application of a finer quantisation, of error resilient source coding, of error resilient channel coding and of combinations of these three. We investigated at different bit errors rates the cases of single-bit errors and of packet loss and burst errors.

In current coding techniques the coding of video objects and their shapes is important. Not many error resilient shape coding techniques exist. We developed a technique which is based on polar co-ordinates and spectral decomposition. Its error resilience is shown, as well as its efficiency in bit rate. The computational load is low.